tempest-react-sdk 0.69.0 → 0.71.0

package/dist/vision/fusion.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"fusion.cjs","names":[],"sources":["../../src/vision/fusion.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * The contract a fused pipeline carries inside its own `.onnx` file.\n *\n * A pipeline built by the Python SDK's `ort_vision_sdk.compose` is a single\n * graph that already contains the detector, the crop-and-resize bridge and the\n * classifier. Everything the runtime needs to drive it — the letterbox\n * resolution, the crop resolution, where the crops come from, how many\n * detections the graph emits, the class names of both stages — was decided at\n * fusion time and written into the model's metadata.\n *\n * This module reads it back. It is the browser half of a contract whose other\n * half lives in `ort_vision_sdk/fusion.py`: the same keys, the same encodings,\n * the same fallbacks. A pipeline fused once therefore runs identically in\n * Python and in a browser tab, off the same file.\n *\n * Building a pipeline stays a Python-side build step — there is no ONNX\n * protobuf writer here, and there is no reason for one: fusing is something you\n * do once next to your export pipeline, not in a page load.\n */\n\nimport { parseNames } from \"./core/metadata\";\n\n/**\n * Which tensor the bridge crops the detected boxes out of.\n *\n * - `\"detector_input\"`: the letterboxed tensor already fed to the detector.\n *   The fused graph then has a **single** image input, but a small object is\n *   cropped out of its downscaled copy.\n * - `\"original\"`: a second, full-resolution image input. The bridge undoes the\n *   letterbox transform in-graph and crops at native resolution. Two tensors to\n *   feed, still one session and one model load.\n */\nexport type CropSource = \"detector_input\" | \"original\";\n\n/** Value of the `ovs.kind` metadata key for a detector→classifier pipeline. */\nexport const FUSION_KIND_DETECT_CLASSIFY = \"detect_classify\";\n\n/**\n * Namespace for every metadata key the fusion writes.\n *\n * Namespaced on purpose: the detector's own Ultralytics metadata (`names`,\n * `task`, `imgsz`) is carried over into the fused model, and an un-prefixed key\n * would either collide with it or be mistaken for it.\n */\nexport const METADATA_PREFIX = \"ovs.\";\n\n/** Name of the fused graph's letterboxed detector input, `[1, 3, H, W]` float32 in `[0, 1]`. */\nexport const INPUT_IMAGE = \"images\";\n\n/** Name of the full-resolution input. Present only when `cropSource === \"original\"`. */\nexport const INPUT_SOURCE = \"source_image\";\n\n/** Name of the `[1]` float32 letterbox scale factor. Only with `cropSource === \"original\"`. */\nexport const INPUT_SCALE = \"letterbox_scale\";\n\n/** Name of the `[2]` float32 `[padLeft, padTop]`. Only with `cropSource === \"original\"`. */\nexport const INPUT_PAD = \"letterbox_pad\";\n\n/**\n * Name of the `[K, 4]` float32 xyxy output, in **letterboxed** input pixels.\n *\n * A file fused by `ort-vision-sdk` 0.9.0 or later reports the box that was\n * actually classified: clamped to the image the crop came from, exactly as\n * RoiAlign received it. Older files report the raw box, so one that ran off the\n * frame draws a rectangle wider than the region the classifier saw.\n */\nexport const OUTPUT_BOXES = \"boxes\";\n\n/** Name of the `[K]` float32 detection-confidence output. */\nexport const OUTPUT_SCORES = \"scores\";\n\n/** Name of the `[K]` int64 detector-class output. */\nexport const OUTPUT_CLASSES = \"classes\";\n\n/** Name of the `[1]` int64 output holding how many of the `K` rows are real. */\nexport const OUTPUT_NUM_DETECTIONS = \"num_detections\";\n\n/** Name of the `[K, numClassifierClasses]` float32 classifier output, one row per box. */\nexport const OUTPUT_PROBS = \"probs\";\n\n/** Everything a fused pipeline declares about how it must be driven. */\nexport interface FusionSpec {\n    /** Pipeline family. Only `\"detect_classify\"` exists today. */\n    readonly kind: string;\n    /** `[width, height]` the detector stage expects — the resolution to letterbox to. */\n    readonly inputSize: readonly [number, number];\n    /** `[width, height]` every crop is resampled to inside the graph. */\n    readonly cropSize: readonly [number, number];\n    /** Which tensor the crops are taken from. */\n    readonly cropSource: CropSource;\n    /**\n     * Fixed number of rows `K` every output carries, surplus zero-padded and\n     * counted by {@link OUTPUT_NUM_DETECTIONS}. `null` means the graph emits\n     * exactly as many rows as survived NMS.\n     */\n    readonly maxDetections: number | null;\n    /** Score threshold baked into the graph's NMS node. */\n    readonly confThreshold: number;\n    /** IoU threshold baked into the graph's NMS node. */\n    readonly iouThreshold: number;\n    /** Whether the classifier stage emits logits that still need a softmax. */\n    readonly applySoftmax: boolean;\n    /** Detector class names in class-id order, or `null` when the fusion recorded none. */\n    readonly detectorNames: readonly string[] | null;\n    /** Classifier class names in class-id order, or `null`. */\n    readonly classifierNames: readonly string[] | null;\n    /** Version of `ort-vision-sdk` that produced the file. */\n    readonly sdkVersion: string;\n    /** Whether driving this pipeline requires feeding the full-resolution input. */\n    readonly needsSourceImage: boolean;\n}\n\nconst KEY_KIND = \"kind\";\nconst KEY_SDK_VERSION = \"sdk_version\";\nconst KEY_INPUT_SIZE = \"input_size\";\nconst KEY_CROP_SIZE = \"crop_size\";\nconst KEY_CROP_SOURCE = \"crop_source\";\nconst KEY_MAX_DETECTIONS = \"max_detections\";\nconst KEY_CONF_THRESHOLD = \"conf_threshold\";\nconst KEY_IOU_THRESHOLD = \"iou_threshold\";\nconst KEY_APPLY_SOFTMAX = \"apply_softmax\";\nconst KEY_DETECTOR_NAMES = \"detector_names\";\nconst KEY_CLASSIFIER_NAMES = \"classifier_names\";\n\nconst DYNAMIC = \"dynamic\";\n\n/**\n * Decode a `\"640,640\"` pair.\n *\n * @param raw The encoded pair.\n * @returns `[width, height]`, or `null` when the value is missing or malformed.\n */\nfunction decodeSize(raw: string | undefined): readonly [number, number] | null {\n    if (!raw) return null;\n    const parts = raw.split(\",\");\n    if (parts.length !== 2) return null;\n    const width = Number(parts[0]);\n    const height = Number(parts[1]);\n    if (!Number.isInteger(width) || !Number.isInteger(height)) return null;\n    if (width < 1 || height < 1) return null;\n    return [width, height];\n}\n\n/**\n * Decode a float, falling back when the value is missing or malformed.\n *\n * @param raw The encoded value.\n * @param fallback Value to use when `raw` cannot be read.\n * @returns The parsed number, or `fallback`.\n */\nfunction decodeFloat(raw: string | undefined, fallback: number): number {\n    if (!raw) return fallback;\n    const value = Number(raw);\n    return Number.isFinite(value) ? value : fallback;\n}\n\n/**\n * Read a pipeline spec out of a model's custom metadata.\n *\n * Individual malformed entries fall back to the value a fusion would have used\n * by default — a single bad float is not a reason to reject an otherwise\n * loadable pipeline. A malformed resolution is fatal, because there is no safe\n * default for one.\n *\n * @param metadata A model's custom metadata map, as read by\n *   {@link readModelMetadata}.\n * @returns The decoded spec, or `null` when the model is not a fused pipeline —\n *   it carries no `ovs.kind` entry, or one naming a pipeline kind this version\n *   does not know how to drive.\n */\nexport function readFusionSpec(\n    metadata: Readonly<Record<string, string>> | undefined,\n): FusionSpec | null {\n    if (!metadata) return null;\n\n    const read: Record<string, string> = {};\n    for (const [key, value] of Object.entries(metadata)) {\n        if (key.startsWith(METADATA_PREFIX)) read[key.slice(METADATA_PREFIX.length)] = value;\n    }\n    if (read[KEY_KIND] !== FUSION_KIND_DETECT_CLASSIFY) return null;\n\n    const inputSize = decodeSize(read[KEY_INPUT_SIZE]);\n    const cropSize = decodeSize(read[KEY_CROP_SIZE]);\n    if (inputSize === null || cropSize === null) return null;\n\n    const rawMax = read[KEY_MAX_DETECTIONS] ?? DYNAMIC;\n    const parsedMax = Number(rawMax);\n    const maxDetections =\n        rawMax === DYNAMIC || !Number.isInteger(parsedMax) || parsedMax < 1 ? null : parsedMax;\n\n    const cropSource: CropSource =\n        read[KEY_CROP_SOURCE] === \"original\" ? \"original\" : \"detector_input\";\n\n    return {\n        kind: FUSION_KIND_DETECT_CLASSIFY,\n        inputSize,\n        cropSize,\n        cropSource,\n        maxDetections,\n        confThreshold: decodeFloat(read[KEY_CONF_THRESHOLD], 0.25),\n        iouThreshold: decodeFloat(read[KEY_IOU_THRESHOLD], 0.45),\n        applySoftmax: (read[KEY_APPLY_SOFTMAX] ?? \"1\") !== \"0\",\n        detectorNames: parseNames(read[KEY_DETECTOR_NAMES]),\n        classifierNames: parseNames(read[KEY_CLASSIFIER_NAMES]),\n        sdkVersion: read[KEY_SDK_VERSION] ?? \"\",\n        needsSourceImage: cropSource === \"original\",\n    };\n}\n"],"mappings":"uCAoCA,IAAa,EAA8B,kBAS9B,EAAkB,OAGlB,EAAc,SAGd,EAAe,eAGf,EAAc,kBAGd,EAAY,gBAUZ,EAAe,QAGf,EAAgB,SAGhB,EAAiB,UAGjB,EAAwB,iBAGxB,EAAe,QAkCtB,EAAW,OACX,EAAkB,cAClB,EAAiB,aACjB,EAAgB,YAChB,EAAkB,cAClB,EAAqB,iBACrB,EAAqB,iBACrB,EAAoB,gBACpB,EAAoB,gBACpB,EAAqB,iBACrB,EAAuB,mBAEvB,EAAU,UAQhB,SAAS,EAAW,EAA2D,CAC3E,GAAI,CAAC,EAAK,OAAO,KACjB,IAAM,EAAQ,EAAI,MAAM,GAAG,EAC3B,GAAI,EAAM,SAAW,EAAG,OAAO,KAC/B,IAAM,EAAQ,OAAO,EAAM,EAAE,EACvB,EAAS,OAAO,EAAM,EAAE,EAG9B,MAFI,CAAC,OAAO,UAAU,CAAK,GAAK,CAAC,OAAO,UAAU,CAAM,GACpD,EAAQ,GAAK,EAAS,EAAU,KAC7B,CAAC,EAAO,CAAM,CACzB,CASA,SAAS,EAAY,EAAyB,EAA0B,CACpE,GAAI,CAAC,EAAK,OAAO,EACjB,IAAM,EAAQ,OAAO,CAAG,EACxB,OAAO,OAAO,SAAS,CAAK,EAAI,EAAQ,CAC5C,CAgBA,SAAgB,EACZ,EACiB,CACjB,GAAI,CAAC,EAAU,OAAO,KAEtB,IAAM,EAA+B,CAAC,EACtC,IAAK,GAAM,CAAC,EAAK,KAAU,OAAO,QAAQ,CAAQ,EAC1C,EAAI,WAAA,MAA0B,IAAG,EAAK,EAAI,MAAM,CAAsB,GAAK,GAEnF,GAAI,EAAK,KAAA,kBAA2C,OAAO,KAE3D,IAAM,EAAY,EAAW,EAAK,EAAe,EAC3C,EAAW,EAAW,EAAK,EAAc,EAC/C,GAAI,IAAc,MAAQ,IAAa,KAAM,OAAO,KAEpD,IAAM,EAAS,EAAK,IAAuB,EACrC,EAAY,OAAO,CAAM,EACzB,EACF,IAAW,GAAW,CAAC,OAAO,UAAU,CAAS,GAAK,EAAY,EAAI,KAAO,EAE3E,EACF,EAAK,KAAqB,WAAa,WAAa,iBAExD,MAAO,CACH,KAAM,EACN,YACA,WACA,aACA,gBACA,cAAe,EAAY,EAAK,GAAqB,GAAI,EACzD,aAAc,EAAY,EAAK,GAAoB,GAAI,EACvD,cAAe,EAAK,IAAsB,OAAS,IACnD,cAAe,EAAA,WAAW,EAAK,EAAmB,EAClD,gBAAiB,EAAA,WAAW,EAAK,EAAqB,EACtD,WAAY,EAAK,IAAoB,GACrC,iBAAkB,IAAe,UACrC,CACJ"}
+{"version":3,"file":"fusion.cjs","names":[],"sources":["../../src/vision/fusion.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * The contract a fused pipeline carries inside its own `.onnx` file.\n *\n * A pipeline built by the Python SDK's `ort_vision_sdk.compose` is a single\n * graph that already contains the detector, the crop-and-resize bridge and the\n * classifier. Everything the runtime needs to drive it — the letterbox\n * resolution, the crop resolution, where the crops come from, how many\n * detections the graph emits, the class names of both stages — was decided at\n * fusion time and written into the model's metadata.\n *\n * This module reads it back. It is the browser half of a contract whose other\n * half lives in `ort_vision_sdk/fusion.py`: the same keys, the same encodings,\n * the same fallbacks. A pipeline fused once therefore runs identically in\n * Python and in a browser tab, off the same file.\n *\n * Building a pipeline stays a Python-side build step — there is no ONNX\n * protobuf writer here, and there is no reason for one: fusing is something you\n * do once next to your export pipeline, not in a page load.\n */\n\nimport { parseNames } from \"./core/metadata\";\n\n/**\n * Which tensor the bridge crops the detected boxes out of.\n *\n * - `\"detector_input\"`: the letterboxed tensor already fed to the detector.\n *   The fused graph then has a **single** image input, but a small object is\n *   cropped out of its downscaled copy.\n * - `\"original\"`: a second, full-resolution image input. The bridge undoes the\n *   letterbox transform in-graph and crops at native resolution. Two tensors to\n *   feed, still one session and one model load.\n */\nexport type CropSource = \"detector_input\" | \"original\";\n\n/** Value of the `ovs.kind` metadata key for a detector→classifier pipeline. */\nexport const FUSION_KIND_DETECT_CLASSIFY = \"detect_classify\";\n\n/**\n * Value of `ovs.kind` for a detector→segmenter→classifier pipeline.\n *\n * The segmenter runs inside each crop rather than over the whole frame: the\n * detector finds the boxes, the bridge cuts them out, and the segmenter turns\n * each crop into a foreground mask that gates the pixels the classifier sees.\n * Fusing one is a build step on the Python side; the browser only loads the\n * result.\n */\nexport const FUSION_KIND_DETECT_SEGMENT_CLASSIFY = \"detect_segment_classify\";\n\n/**\n * Namespace for every metadata key the fusion writes.\n *\n * Namespaced on purpose: the detector's own Ultralytics metadata (`names`,\n * `task`, `imgsz`) is carried over into the fused model, and an un-prefixed key\n * would either collide with it or be mistaken for it.\n */\nexport const METADATA_PREFIX = \"ovs.\";\n\n/** Name of the fused graph's letterboxed detector input, `[1, 3, H, W]` float32 in `[0, 1]`. */\nexport const INPUT_IMAGE = \"images\";\n\n/** Name of the full-resolution input. Present only when `cropSource === \"original\"`. */\nexport const INPUT_SOURCE = \"source_image\";\n\n/** Name of the `[1]` float32 letterbox scale factor. Only with `cropSource === \"original\"`. */\nexport const INPUT_SCALE = \"letterbox_scale\";\n\n/** Name of the `[2]` float32 `[padLeft, padTop]`. Only with `cropSource === \"original\"`. */\nexport const INPUT_PAD = \"letterbox_pad\";\n\n/**\n * Name of the `[K, 4]` float32 xyxy output, in **letterboxed** input pixels.\n *\n * A file fused by `ort-vision-sdk` 0.9.0 or later reports the box that was\n * actually classified: clamped to the image the crop came from, exactly as\n * RoiAlign received it. Older files report the raw box, so one that ran off the\n * frame draws a rectangle wider than the region the classifier saw.\n */\nexport const OUTPUT_BOXES = \"boxes\";\n\n/** Name of the `[K]` float32 detection-confidence output. */\nexport const OUTPUT_SCORES = \"scores\";\n\n/** Name of the `[K]` int64 detector-class output. */\nexport const OUTPUT_CLASSES = \"classes\";\n\n/** Name of the `[1]` int64 output holding how many of the `K` rows are real. */\nexport const OUTPUT_NUM_DETECTIONS = \"num_detections\";\n\n/** Name of the `[K, numClassifierClasses]` float32 classifier output, one row per box. */\n/**\n * Name of the `(K, 1, cropH, cropW)` float32 mask output. Only on a\n * `detect_segment_classify` pipeline.\n *\n * The mask lives in the **crop's** coordinate space, not the original image's,\n * because that is where the segmenter computed it. Values are 0 or 1 after\n * thresholding; padded rows are all-zero, like every other output.\n */\nexport const OUTPUT_MASKS = \"masks\";\n\nexport const OUTPUT_PROBS = \"probs\";\n\n/** Everything a fused pipeline declares about how it must be driven. */\nexport interface FusionSpec {\n    /** Pipeline family — {@link FUSION_KIND_DETECT_CLASSIFY} or {@link FUSION_KIND_DETECT_SEGMENT_CLASSIFY}. */\n    readonly kind: string;\n    /** `[width, height]` the detector stage expects — the resolution to letterbox to. */\n    readonly inputSize: readonly [number, number];\n    /** `[width, height]` every crop is resampled to inside the graph. */\n    readonly cropSize: readonly [number, number];\n    /** Which tensor the crops are taken from. */\n    readonly cropSource: CropSource;\n    /**\n     * Fixed number of rows `K` every output carries, surplus zero-padded and\n     * counted by {@link OUTPUT_NUM_DETECTIONS}. `null` means the graph emits\n     * exactly as many rows as survived NMS.\n     */\n    readonly maxDetections: number | null;\n    /** Score threshold baked into the graph's NMS node. */\n    readonly confThreshold: number;\n    /** IoU threshold baked into the graph's NMS node. */\n    readonly iouThreshold: number;\n    /** Whether the classifier stage emits logits that still need a softmax. */\n    readonly applySoftmax: boolean;\n    /** Detector class names in class-id order, or `null` when the fusion recorded none. */\n    readonly detectorNames: readonly string[] | null;\n    /** Classifier class names in class-id order, or `null`. */\n    readonly classifierNames: readonly string[] | null;\n    /** Version of `ort-vision-sdk` that produced the file. */\n    readonly sdkVersion: string;\n    /** Whether driving this pipeline requires feeding the full-resolution input. */\n    readonly needsSourceImage: boolean;\n    /** Whether this pipeline emits a {@link OUTPUT_MASKS} tensor. */\n    readonly hasMasks: boolean;\n    /** Class id → name for the segmenter's channels, or `null`. */\n    readonly segmenterNames: readonly string[] | null;\n    /**\n     * Probability above which a segmenter pixel counted as foreground.\n     *\n     * Baked into the graph at fusion time; reported so a caller can say what the\n     * masks mean.\n     */\n    readonly maskThreshold: number;\n    /** Whether the mask gated the crop before the classifier saw it. */\n    readonly maskApplied: boolean;\n}\n\nconst KEY_KIND = \"kind\";\nconst KEY_SDK_VERSION = \"sdk_version\";\nconst KEY_INPUT_SIZE = \"input_size\";\nconst KEY_CROP_SIZE = \"crop_size\";\nconst KEY_CROP_SOURCE = \"crop_source\";\nconst KEY_MAX_DETECTIONS = \"max_detections\";\nconst KEY_CONF_THRESHOLD = \"conf_threshold\";\nconst KEY_IOU_THRESHOLD = \"iou_threshold\";\nconst KEY_APPLY_SOFTMAX = \"apply_softmax\";\nconst KEY_DETECTOR_NAMES = \"detector_names\";\nconst KEY_CLASSIFIER_NAMES = \"classifier_names\";\nconst KEY_SEGMENTER_NAMES = \"segmenter_names\";\nconst KEY_MASK_THRESHOLD = \"mask_threshold\";\nconst KEY_MASK_APPLIED = \"mask_applied\";\n\nconst DYNAMIC = \"dynamic\";\n\n/**\n * Decode a `\"640,640\"` pair.\n *\n * @param raw The encoded pair.\n * @returns `[width, height]`, or `null` when the value is missing or malformed.\n */\nfunction decodeSize(raw: string | undefined): readonly [number, number] | null {\n    if (!raw) return null;\n    const parts = raw.split(\",\");\n    if (parts.length !== 2) return null;\n    const width = Number(parts[0]);\n    const height = Number(parts[1]);\n    if (!Number.isInteger(width) || !Number.isInteger(height)) return null;\n    if (width < 1 || height < 1) return null;\n    return [width, height];\n}\n\n/**\n * Decode a float, falling back when the value is missing or malformed.\n *\n * @param raw The encoded value.\n * @param fallback Value to use when `raw` cannot be read.\n * @returns The parsed number, or `fallback`.\n */\nfunction decodeFloat(raw: string | undefined, fallback: number): number {\n    if (!raw) return fallback;\n    const value = Number(raw);\n    return Number.isFinite(value) ? value : fallback;\n}\n\n/**\n * Read a pipeline spec out of a model's custom metadata.\n *\n * Individual malformed entries fall back to the value a fusion would have used\n * by default — a single bad float is not a reason to reject an otherwise\n * loadable pipeline. A malformed resolution is fatal, because there is no safe\n * default for one.\n *\n * @param metadata A model's custom metadata map, as read by\n *   {@link readModelMetadata}.\n * @returns The decoded spec, or `null` when the model is not a fused pipeline —\n *   it carries no `ovs.kind` entry, or one naming a pipeline kind this version\n *   does not know how to drive.\n */\nexport function readFusionSpec(\n    metadata: Readonly<Record<string, string>> | undefined,\n): FusionSpec | null {\n    if (!metadata) return null;\n\n    const read: Record<string, string> = {};\n    for (const [key, value] of Object.entries(metadata)) {\n        if (key.startsWith(METADATA_PREFIX)) read[key.slice(METADATA_PREFIX.length)] = value;\n    }\n    const kind = read[KEY_KIND];\n    if (kind !== FUSION_KIND_DETECT_CLASSIFY && kind !== FUSION_KIND_DETECT_SEGMENT_CLASSIFY) {\n        return null;\n    }\n\n    const inputSize = decodeSize(read[KEY_INPUT_SIZE]);\n    const cropSize = decodeSize(read[KEY_CROP_SIZE]);\n    if (inputSize === null || cropSize === null) return null;\n\n    const rawMax = read[KEY_MAX_DETECTIONS] ?? DYNAMIC;\n    const parsedMax = Number(rawMax);\n    const maxDetections =\n        rawMax === DYNAMIC || !Number.isInteger(parsedMax) || parsedMax < 1 ? null : parsedMax;\n\n    const cropSource: CropSource =\n        read[KEY_CROP_SOURCE] === \"original\" ? \"original\" : \"detector_input\";\n\n    return {\n        kind,\n        inputSize,\n        cropSize,\n        cropSource,\n        maxDetections,\n        confThreshold: decodeFloat(read[KEY_CONF_THRESHOLD], 0.25),\n        iouThreshold: decodeFloat(read[KEY_IOU_THRESHOLD], 0.45),\n        applySoftmax: (read[KEY_APPLY_SOFTMAX] ?? \"1\") !== \"0\",\n        detectorNames: parseNames(read[KEY_DETECTOR_NAMES]),\n        classifierNames: parseNames(read[KEY_CLASSIFIER_NAMES]),\n        sdkVersion: read[KEY_SDK_VERSION] ?? \"\",\n        needsSourceImage: cropSource === \"original\",\n        hasMasks: kind === FUSION_KIND_DETECT_SEGMENT_CLASSIFY,\n        segmenterNames: parseNames(read[KEY_SEGMENTER_NAMES]),\n        maskThreshold: decodeFloat(read[KEY_MASK_THRESHOLD], 0.5),\n        maskApplied: (read[KEY_MASK_APPLIED] ?? \"1\") !== \"0\",\n    };\n}\n"],"mappings":"uCAoCA,IAAa,EAA8B,kBAW9B,EAAsC,0BAStC,EAAkB,OAGlB,EAAc,SAGd,EAAe,eAGf,EAAc,kBAGd,EAAY,gBAUZ,EAAe,QAGf,EAAgB,SAGhB,EAAiB,UAGjB,EAAwB,iBAWxB,EAAe,QAEf,EAAe,QA+CtB,EAAW,OACX,EAAkB,cAClB,EAAiB,aACjB,EAAgB,YAChB,EAAkB,cAClB,EAAqB,iBACrB,EAAqB,iBACrB,EAAoB,gBACpB,EAAoB,gBACpB,EAAqB,iBACrB,EAAuB,mBACvB,EAAsB,kBACtB,EAAqB,iBACrB,EAAmB,eAEnB,EAAU,UAQhB,SAAS,EAAW,EAA2D,CAC3E,GAAI,CAAC,EAAK,OAAO,KACjB,IAAM,EAAQ,EAAI,MAAM,GAAG,EAC3B,GAAI,EAAM,SAAW,EAAG,OAAO,KAC/B,IAAM,EAAQ,OAAO,EAAM,EAAE,EACvB,EAAS,OAAO,EAAM,EAAE,EAG9B,MAFI,CAAC,OAAO,UAAU,CAAK,GAAK,CAAC,OAAO,UAAU,CAAM,GACpD,EAAQ,GAAK,EAAS,EAAU,KAC7B,CAAC,EAAO,CAAM,CACzB,CASA,SAAS,EAAY,EAAyB,EAA0B,CACpE,GAAI,CAAC,EAAK,OAAO,EACjB,IAAM,EAAQ,OAAO,CAAG,EACxB,OAAO,OAAO,SAAS,CAAK,EAAI,EAAQ,CAC5C,CAgBA,SAAgB,EACZ,EACiB,CACjB,GAAI,CAAC,EAAU,OAAO,KAEtB,IAAM,EAA+B,CAAC,EACtC,IAAK,GAAM,CAAC,EAAK,KAAU,OAAO,QAAQ,CAAQ,EAC1C,EAAI,WAAA,MAA0B,IAAG,EAAK,EAAI,MAAM,CAAsB,GAAK,GAEnF,IAAM,EAAO,EAAK,GAClB,GAAI,IAAA,mBAAwC,IAAA,0BACxC,OAAO,KAGX,IAAM,EAAY,EAAW,EAAK,EAAe,EAC3C,EAAW,EAAW,EAAK,EAAc,EAC/C,GAAI,IAAc,MAAQ,IAAa,KAAM,OAAO,KAEpD,IAAM,EAAS,EAAK,IAAuB,EACrC,EAAY,OAAO,CAAM,EACzB,EACF,IAAW,GAAW,CAAC,OAAO,UAAU,CAAS,GAAK,EAAY,EAAI,KAAO,EAE3E,EACF,EAAK,KAAqB,WAAa,WAAa,iBAExD,MAAO,CACH,OACA,YACA,WACA,aACA,gBACA,cAAe,EAAY,EAAK,GAAqB,GAAI,EACzD,aAAc,EAAY,EAAK,GAAoB,GAAI,EACvD,cAAe,EAAK,IAAsB,OAAS,IACnD,cAAe,EAAA,WAAW,EAAK,EAAmB,EAClD,gBAAiB,EAAA,WAAW,EAAK,EAAqB,EACtD,WAAY,EAAK,IAAoB,GACrC,iBAAkB,IAAe,WACjC,SAAU,IAAS,EACnB,eAAgB,EAAA,WAAW,EAAK,EAAoB,EACpD,cAAe,EAAY,EAAK,GAAqB,EAAG,EACxD,aAAc,EAAK,IAAqB,OAAS,GACrD,CACJ"}

package/dist/vision/fusion.js

@@ -1,42 +1,47 @@
 import { parseNames as e } from "./core/metadata.js";
 //#region src/vision/fusion.ts
-var t = "detect_classify", n = "ovs.", r = "images", i = "source_image", a = "letterbox_scale", o = "letterbox_pad", s = "boxes", c = "scores", l = "classes", u = "num_detections", d = "probs", f = "kind", p = "sdk_version", m = "input_size", h = "crop_size", g = "crop_source", _ = "max_detections", v = "conf_threshold", y = "iou_threshold", b = "apply_softmax", x = "detector_names", S = "classifier_names", C = "dynamic";
-function w(e) {
+var t = "detect_classify", n = "detect_segment_classify", r = "ovs.", i = "images", a = "source_image", o = "letterbox_scale", s = "letterbox_pad", c = "boxes", l = "scores", u = "classes", d = "num_detections", f = "masks", p = "probs", m = "kind", h = "sdk_version", g = "input_size", _ = "crop_size", v = "crop_source", y = "max_detections", b = "conf_threshold", x = "iou_threshold", S = "apply_softmax", C = "detector_names", w = "classifier_names", T = "segmenter_names", E = "mask_threshold", D = "mask_applied", O = "dynamic";
+function k(e) {
 	if (!e) return null;
 	let t = e.split(",");
 	if (t.length !== 2) return null;
 	let n = Number(t[0]), r = Number(t[1]);
 	return !Number.isInteger(n) || !Number.isInteger(r) || n < 1 || r < 1 ? null : [n, r];
 }
-function T(e, t) {
+function A(e, t) {
 	if (!e) return t;
 	let n = Number(e);
 	return Number.isFinite(n) ? n : t;
 }
-function E(n) {
-	if (!n) return null;
+function j(t) {
+	if (!t) return null;
 	let r = {};
-	for (let [e, t] of Object.entries(n)) e.startsWith("ovs.") && (r[e.slice(4)] = t);
-	if (r[f] !== "detect_classify") return null;
-	let i = w(r[m]), a = w(r[h]);
-	if (i === null || a === null) return null;
-	let o = r[_] ?? C, s = Number(o), c = o === C || !Number.isInteger(s) || s < 1 ? null : s, l = r[g] === "original" ? "original" : "detector_input";
+	for (let [e, n] of Object.entries(t)) e.startsWith("ovs.") && (r[e.slice(4)] = n);
+	let i = r[m];
+	if (i !== "detect_classify" && i !== "detect_segment_classify") return null;
+	let a = k(r[g]), o = k(r[_]);
+	if (a === null || o === null) return null;
+	let s = r[y] ?? O, c = Number(s), l = s === O || !Number.isInteger(c) || c < 1 ? null : c, u = r[v] === "original" ? "original" : "detector_input";
 	return {
-		kind: t,
-		inputSize: i,
-		cropSize: a,
-		cropSource: l,
-		maxDetections: c,
-		confThreshold: T(r[v], .25),
-		iouThreshold: T(r[y], .45),
-		applySoftmax: (r[b] ?? "1") !== "0",
-		detectorNames: e(r[x]),
-		classifierNames: e(r[S]),
-		sdkVersion: r[p] ?? "",
-		needsSourceImage: l === "original"
+		kind: i,
+		inputSize: a,
+		cropSize: o,
+		cropSource: u,
+		maxDetections: l,
+		confThreshold: A(r[b], .25),
+		iouThreshold: A(r[x], .45),
+		applySoftmax: (r[S] ?? "1") !== "0",
+		detectorNames: e(r[C]),
+		classifierNames: e(r[w]),
+		sdkVersion: r[h] ?? "",
+		needsSourceImage: u === "original",
+		hasMasks: i === n,
+		segmenterNames: e(r[T]),
+		maskThreshold: A(r[E], .5),
+		maskApplied: (r[D] ?? "1") !== "0"
 	};
 }
 //#endregion
-export { t as FUSION_KIND_DETECT_CLASSIFY, r as INPUT_IMAGE, o as INPUT_PAD, a as INPUT_SCALE, i as INPUT_SOURCE, n as METADATA_PREFIX, s as OUTPUT_BOXES, l as OUTPUT_CLASSES, u as OUTPUT_NUM_DETECTIONS, d as OUTPUT_PROBS, c as OUTPUT_SCORES, E as readFusionSpec };
+export { t as FUSION_KIND_DETECT_CLASSIFY, n as FUSION_KIND_DETECT_SEGMENT_CLASSIFY, i as INPUT_IMAGE, s as INPUT_PAD, o as INPUT_SCALE, a as INPUT_SOURCE, r as METADATA_PREFIX, c as OUTPUT_BOXES, u as OUTPUT_CLASSES, f as OUTPUT_MASKS, d as OUTPUT_NUM_DETECTIONS, p as OUTPUT_PROBS, l as OUTPUT_SCORES, j as readFusionSpec };
 
 //# sourceMappingURL=fusion.js.map

package/dist/vision/fusion.js.map

@@ -1 +1 @@
-{"version":3,"file":"fusion.js","names":[],"sources":["../../src/vision/fusion.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * The contract a fused pipeline carries inside its own `.onnx` file.\n *\n * A pipeline built by the Python SDK's `ort_vision_sdk.compose` is a single\n * graph that already contains the detector, the crop-and-resize bridge and the\n * classifier. Everything the runtime needs to drive it — the letterbox\n * resolution, the crop resolution, where the crops come from, how many\n * detections the graph emits, the class names of both stages — was decided at\n * fusion time and written into the model's metadata.\n *\n * This module reads it back. It is the browser half of a contract whose other\n * half lives in `ort_vision_sdk/fusion.py`: the same keys, the same encodings,\n * the same fallbacks. A pipeline fused once therefore runs identically in\n * Python and in a browser tab, off the same file.\n *\n * Building a pipeline stays a Python-side build step — there is no ONNX\n * protobuf writer here, and there is no reason for one: fusing is something you\n * do once next to your export pipeline, not in a page load.\n */\n\nimport { parseNames } from \"./core/metadata\";\n\n/**\n * Which tensor the bridge crops the detected boxes out of.\n *\n * - `\"detector_input\"`: the letterboxed tensor already fed to the detector.\n *   The fused graph then has a **single** image input, but a small object is\n *   cropped out of its downscaled copy.\n * - `\"original\"`: a second, full-resolution image input. The bridge undoes the\n *   letterbox transform in-graph and crops at native resolution. Two tensors to\n *   feed, still one session and one model load.\n */\nexport type CropSource = \"detector_input\" | \"original\";\n\n/** Value of the `ovs.kind` metadata key for a detector→classifier pipeline. */\nexport const FUSION_KIND_DETECT_CLASSIFY = \"detect_classify\";\n\n/**\n * Namespace for every metadata key the fusion writes.\n *\n * Namespaced on purpose: the detector's own Ultralytics metadata (`names`,\n * `task`, `imgsz`) is carried over into the fused model, and an un-prefixed key\n * would either collide with it or be mistaken for it.\n */\nexport const METADATA_PREFIX = \"ovs.\";\n\n/** Name of the fused graph's letterboxed detector input, `[1, 3, H, W]` float32 in `[0, 1]`. */\nexport const INPUT_IMAGE = \"images\";\n\n/** Name of the full-resolution input. Present only when `cropSource === \"original\"`. */\nexport const INPUT_SOURCE = \"source_image\";\n\n/** Name of the `[1]` float32 letterbox scale factor. Only with `cropSource === \"original\"`. */\nexport const INPUT_SCALE = \"letterbox_scale\";\n\n/** Name of the `[2]` float32 `[padLeft, padTop]`. Only with `cropSource === \"original\"`. */\nexport const INPUT_PAD = \"letterbox_pad\";\n\n/**\n * Name of the `[K, 4]` float32 xyxy output, in **letterboxed** input pixels.\n *\n * A file fused by `ort-vision-sdk` 0.9.0 or later reports the box that was\n * actually classified: clamped to the image the crop came from, exactly as\n * RoiAlign received it. Older files report the raw box, so one that ran off the\n * frame draws a rectangle wider than the region the classifier saw.\n */\nexport const OUTPUT_BOXES = \"boxes\";\n\n/** Name of the `[K]` float32 detection-confidence output. */\nexport const OUTPUT_SCORES = \"scores\";\n\n/** Name of the `[K]` int64 detector-class output. */\nexport const OUTPUT_CLASSES = \"classes\";\n\n/** Name of the `[1]` int64 output holding how many of the `K` rows are real. */\nexport const OUTPUT_NUM_DETECTIONS = \"num_detections\";\n\n/** Name of the `[K, numClassifierClasses]` float32 classifier output, one row per box. */\nexport const OUTPUT_PROBS = \"probs\";\n\n/** Everything a fused pipeline declares about how it must be driven. */\nexport interface FusionSpec {\n    /** Pipeline family. Only `\"detect_classify\"` exists today. */\n    readonly kind: string;\n    /** `[width, height]` the detector stage expects — the resolution to letterbox to. */\n    readonly inputSize: readonly [number, number];\n    /** `[width, height]` every crop is resampled to inside the graph. */\n    readonly cropSize: readonly [number, number];\n    /** Which tensor the crops are taken from. */\n    readonly cropSource: CropSource;\n    /**\n     * Fixed number of rows `K` every output carries, surplus zero-padded and\n     * counted by {@link OUTPUT_NUM_DETECTIONS}. `null` means the graph emits\n     * exactly as many rows as survived NMS.\n     */\n    readonly maxDetections: number | null;\n    /** Score threshold baked into the graph's NMS node. */\n    readonly confThreshold: number;\n    /** IoU threshold baked into the graph's NMS node. */\n    readonly iouThreshold: number;\n    /** Whether the classifier stage emits logits that still need a softmax. */\n    readonly applySoftmax: boolean;\n    /** Detector class names in class-id order, or `null` when the fusion recorded none. */\n    readonly detectorNames: readonly string[] | null;\n    /** Classifier class names in class-id order, or `null`. */\n    readonly classifierNames: readonly string[] | null;\n    /** Version of `ort-vision-sdk` that produced the file. */\n    readonly sdkVersion: string;\n    /** Whether driving this pipeline requires feeding the full-resolution input. */\n    readonly needsSourceImage: boolean;\n}\n\nconst KEY_KIND = \"kind\";\nconst KEY_SDK_VERSION = \"sdk_version\";\nconst KEY_INPUT_SIZE = \"input_size\";\nconst KEY_CROP_SIZE = \"crop_size\";\nconst KEY_CROP_SOURCE = \"crop_source\";\nconst KEY_MAX_DETECTIONS = \"max_detections\";\nconst KEY_CONF_THRESHOLD = \"conf_threshold\";\nconst KEY_IOU_THRESHOLD = \"iou_threshold\";\nconst KEY_APPLY_SOFTMAX = \"apply_softmax\";\nconst KEY_DETECTOR_NAMES = \"detector_names\";\nconst KEY_CLASSIFIER_NAMES = \"classifier_names\";\n\nconst DYNAMIC = \"dynamic\";\n\n/**\n * Decode a `\"640,640\"` pair.\n *\n * @param raw The encoded pair.\n * @returns `[width, height]`, or `null` when the value is missing or malformed.\n */\nfunction decodeSize(raw: string | undefined): readonly [number, number] | null {\n    if (!raw) return null;\n    const parts = raw.split(\",\");\n    if (parts.length !== 2) return null;\n    const width = Number(parts[0]);\n    const height = Number(parts[1]);\n    if (!Number.isInteger(width) || !Number.isInteger(height)) return null;\n    if (width < 1 || height < 1) return null;\n    return [width, height];\n}\n\n/**\n * Decode a float, falling back when the value is missing or malformed.\n *\n * @param raw The encoded value.\n * @param fallback Value to use when `raw` cannot be read.\n * @returns The parsed number, or `fallback`.\n */\nfunction decodeFloat(raw: string | undefined, fallback: number): number {\n    if (!raw) return fallback;\n    const value = Number(raw);\n    return Number.isFinite(value) ? value : fallback;\n}\n\n/**\n * Read a pipeline spec out of a model's custom metadata.\n *\n * Individual malformed entries fall back to the value a fusion would have used\n * by default — a single bad float is not a reason to reject an otherwise\n * loadable pipeline. A malformed resolution is fatal, because there is no safe\n * default for one.\n *\n * @param metadata A model's custom metadata map, as read by\n *   {@link readModelMetadata}.\n * @returns The decoded spec, or `null` when the model is not a fused pipeline —\n *   it carries no `ovs.kind` entry, or one naming a pipeline kind this version\n *   does not know how to drive.\n */\nexport function readFusionSpec(\n    metadata: Readonly<Record<string, string>> | undefined,\n): FusionSpec | null {\n    if (!metadata) return null;\n\n    const read: Record<string, string> = {};\n    for (const [key, value] of Object.entries(metadata)) {\n        if (key.startsWith(METADATA_PREFIX)) read[key.slice(METADATA_PREFIX.length)] = value;\n    }\n    if (read[KEY_KIND] !== FUSION_KIND_DETECT_CLASSIFY) return null;\n\n    const inputSize = decodeSize(read[KEY_INPUT_SIZE]);\n    const cropSize = decodeSize(read[KEY_CROP_SIZE]);\n    if (inputSize === null || cropSize === null) return null;\n\n    const rawMax = read[KEY_MAX_DETECTIONS] ?? DYNAMIC;\n    const parsedMax = Number(rawMax);\n    const maxDetections =\n        rawMax === DYNAMIC || !Number.isInteger(parsedMax) || parsedMax < 1 ? null : parsedMax;\n\n    const cropSource: CropSource =\n        read[KEY_CROP_SOURCE] === \"original\" ? \"original\" : \"detector_input\";\n\n    return {\n        kind: FUSION_KIND_DETECT_CLASSIFY,\n        inputSize,\n        cropSize,\n        cropSource,\n        maxDetections,\n        confThreshold: decodeFloat(read[KEY_CONF_THRESHOLD], 0.25),\n        iouThreshold: decodeFloat(read[KEY_IOU_THRESHOLD], 0.45),\n        applySoftmax: (read[KEY_APPLY_SOFTMAX] ?? \"1\") !== \"0\",\n        detectorNames: parseNames(read[KEY_DETECTOR_NAMES]),\n        classifierNames: parseNames(read[KEY_CLASSIFIER_NAMES]),\n        sdkVersion: read[KEY_SDK_VERSION] ?? \"\",\n        needsSourceImage: cropSource === \"original\",\n    };\n}\n"],"mappings":";;AAoCA,IAAa,IAA8B,mBAS9B,IAAkB,QAGlB,IAAc,UAGd,IAAe,gBAGf,IAAc,mBAGd,IAAY,iBAUZ,IAAe,SAGf,IAAgB,UAGhB,IAAiB,WAGjB,IAAwB,kBAGxB,IAAe,SAkCtB,IAAW,QACX,IAAkB,eAClB,IAAiB,cACjB,IAAgB,aAChB,IAAkB,eAClB,IAAqB,kBACrB,IAAqB,kBACrB,IAAoB,iBACpB,IAAoB,iBACpB,IAAqB,kBACrB,IAAuB,oBAEvB,IAAU;AAQhB,SAAS,EAAW,GAA2D;CAC3E,IAAI,CAAC,GAAK,OAAO;CACjB,IAAM,IAAQ,EAAI,MAAM,GAAG;CAC3B,IAAI,EAAM,WAAW,GAAG,OAAO;CAC/B,IAAM,IAAQ,OAAO,EAAM,EAAE,GACvB,IAAS,OAAO,EAAM,EAAE;CAG9B,OAFI,CAAC,OAAO,UAAU,CAAK,KAAK,CAAC,OAAO,UAAU,CAAM,KACpD,IAAQ,KAAK,IAAS,IAAU,OAC7B,CAAC,GAAO,CAAM;AACzB;AASA,SAAS,EAAY,GAAyB,GAA0B;CACpE,IAAI,CAAC,GAAK,OAAO;CACjB,IAAM,IAAQ,OAAO,CAAG;CACxB,OAAO,OAAO,SAAS,CAAK,IAAI,IAAQ;AAC5C;AAgBA,SAAgB,EACZ,GACiB;CACjB,IAAI,CAAC,GAAU,OAAO;CAEtB,IAAM,IAA+B,CAAC;CACtC,KAAK,IAAM,CAAC,GAAK,MAAU,OAAO,QAAQ,CAAQ,GAC9C,AAAI,EAAI,WAAA,MAA0B,MAAG,EAAK,EAAI,MAAM,CAAsB,KAAK;CAEnF,IAAI,EAAK,OAAA,mBAA2C,OAAO;CAE3D,IAAM,IAAY,EAAW,EAAK,EAAe,GAC3C,IAAW,EAAW,EAAK,EAAc;CAC/C,IAAI,MAAc,QAAQ,MAAa,MAAM,OAAO;CAEpD,IAAM,IAAS,EAAK,MAAuB,GACrC,IAAY,OAAO,CAAM,GACzB,IACF,MAAW,KAAW,CAAC,OAAO,UAAU,CAAS,KAAK,IAAY,IAAI,OAAO,GAE3E,IACF,EAAK,OAAqB,aAAa,aAAa;CAExD,OAAO;EACH,MAAM;EACN;EACA;EACA;EACA;EACA,eAAe,EAAY,EAAK,IAAqB,GAAI;EACzD,cAAc,EAAY,EAAK,IAAoB,GAAI;EACvD,eAAe,EAAK,MAAsB,SAAS;EACnD,eAAe,EAAW,EAAK,EAAmB;EAClD,iBAAiB,EAAW,EAAK,EAAqB;EACtD,YAAY,EAAK,MAAoB;EACrC,kBAAkB,MAAe;CACrC;AACJ"}
+{"version":3,"file":"fusion.js","names":[],"sources":["../../src/vision/fusion.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * The contract a fused pipeline carries inside its own `.onnx` file.\n *\n * A pipeline built by the Python SDK's `ort_vision_sdk.compose` is a single\n * graph that already contains the detector, the crop-and-resize bridge and the\n * classifier. Everything the runtime needs to drive it — the letterbox\n * resolution, the crop resolution, where the crops come from, how many\n * detections the graph emits, the class names of both stages — was decided at\n * fusion time and written into the model's metadata.\n *\n * This module reads it back. It is the browser half of a contract whose other\n * half lives in `ort_vision_sdk/fusion.py`: the same keys, the same encodings,\n * the same fallbacks. A pipeline fused once therefore runs identically in\n * Python and in a browser tab, off the same file.\n *\n * Building a pipeline stays a Python-side build step — there is no ONNX\n * protobuf writer here, and there is no reason for one: fusing is something you\n * do once next to your export pipeline, not in a page load.\n */\n\nimport { parseNames } from \"./core/metadata\";\n\n/**\n * Which tensor the bridge crops the detected boxes out of.\n *\n * - `\"detector_input\"`: the letterboxed tensor already fed to the detector.\n *   The fused graph then has a **single** image input, but a small object is\n *   cropped out of its downscaled copy.\n * - `\"original\"`: a second, full-resolution image input. The bridge undoes the\n *   letterbox transform in-graph and crops at native resolution. Two tensors to\n *   feed, still one session and one model load.\n */\nexport type CropSource = \"detector_input\" | \"original\";\n\n/** Value of the `ovs.kind` metadata key for a detector→classifier pipeline. */\nexport const FUSION_KIND_DETECT_CLASSIFY = \"detect_classify\";\n\n/**\n * Value of `ovs.kind` for a detector→segmenter→classifier pipeline.\n *\n * The segmenter runs inside each crop rather than over the whole frame: the\n * detector finds the boxes, the bridge cuts them out, and the segmenter turns\n * each crop into a foreground mask that gates the pixels the classifier sees.\n * Fusing one is a build step on the Python side; the browser only loads the\n * result.\n */\nexport const FUSION_KIND_DETECT_SEGMENT_CLASSIFY = \"detect_segment_classify\";\n\n/**\n * Namespace for every metadata key the fusion writes.\n *\n * Namespaced on purpose: the detector's own Ultralytics metadata (`names`,\n * `task`, `imgsz`) is carried over into the fused model, and an un-prefixed key\n * would either collide with it or be mistaken for it.\n */\nexport const METADATA_PREFIX = \"ovs.\";\n\n/** Name of the fused graph's letterboxed detector input, `[1, 3, H, W]` float32 in `[0, 1]`. */\nexport const INPUT_IMAGE = \"images\";\n\n/** Name of the full-resolution input. Present only when `cropSource === \"original\"`. */\nexport const INPUT_SOURCE = \"source_image\";\n\n/** Name of the `[1]` float32 letterbox scale factor. Only with `cropSource === \"original\"`. */\nexport const INPUT_SCALE = \"letterbox_scale\";\n\n/** Name of the `[2]` float32 `[padLeft, padTop]`. Only with `cropSource === \"original\"`. */\nexport const INPUT_PAD = \"letterbox_pad\";\n\n/**\n * Name of the `[K, 4]` float32 xyxy output, in **letterboxed** input pixels.\n *\n * A file fused by `ort-vision-sdk` 0.9.0 or later reports the box that was\n * actually classified: clamped to the image the crop came from, exactly as\n * RoiAlign received it. Older files report the raw box, so one that ran off the\n * frame draws a rectangle wider than the region the classifier saw.\n */\nexport const OUTPUT_BOXES = \"boxes\";\n\n/** Name of the `[K]` float32 detection-confidence output. */\nexport const OUTPUT_SCORES = \"scores\";\n\n/** Name of the `[K]` int64 detector-class output. */\nexport const OUTPUT_CLASSES = \"classes\";\n\n/** Name of the `[1]` int64 output holding how many of the `K` rows are real. */\nexport const OUTPUT_NUM_DETECTIONS = \"num_detections\";\n\n/** Name of the `[K, numClassifierClasses]` float32 classifier output, one row per box. */\n/**\n * Name of the `(K, 1, cropH, cropW)` float32 mask output. Only on a\n * `detect_segment_classify` pipeline.\n *\n * The mask lives in the **crop's** coordinate space, not the original image's,\n * because that is where the segmenter computed it. Values are 0 or 1 after\n * thresholding; padded rows are all-zero, like every other output.\n */\nexport const OUTPUT_MASKS = \"masks\";\n\nexport const OUTPUT_PROBS = \"probs\";\n\n/** Everything a fused pipeline declares about how it must be driven. */\nexport interface FusionSpec {\n    /** Pipeline family — {@link FUSION_KIND_DETECT_CLASSIFY} or {@link FUSION_KIND_DETECT_SEGMENT_CLASSIFY}. */\n    readonly kind: string;\n    /** `[width, height]` the detector stage expects — the resolution to letterbox to. */\n    readonly inputSize: readonly [number, number];\n    /** `[width, height]` every crop is resampled to inside the graph. */\n    readonly cropSize: readonly [number, number];\n    /** Which tensor the crops are taken from. */\n    readonly cropSource: CropSource;\n    /**\n     * Fixed number of rows `K` every output carries, surplus zero-padded and\n     * counted by {@link OUTPUT_NUM_DETECTIONS}. `null` means the graph emits\n     * exactly as many rows as survived NMS.\n     */\n    readonly maxDetections: number | null;\n    /** Score threshold baked into the graph's NMS node. */\n    readonly confThreshold: number;\n    /** IoU threshold baked into the graph's NMS node. */\n    readonly iouThreshold: number;\n    /** Whether the classifier stage emits logits that still need a softmax. */\n    readonly applySoftmax: boolean;\n    /** Detector class names in class-id order, or `null` when the fusion recorded none. */\n    readonly detectorNames: readonly string[] | null;\n    /** Classifier class names in class-id order, or `null`. */\n    readonly classifierNames: readonly string[] | null;\n    /** Version of `ort-vision-sdk` that produced the file. */\n    readonly sdkVersion: string;\n    /** Whether driving this pipeline requires feeding the full-resolution input. */\n    readonly needsSourceImage: boolean;\n    /** Whether this pipeline emits a {@link OUTPUT_MASKS} tensor. */\n    readonly hasMasks: boolean;\n    /** Class id → name for the segmenter's channels, or `null`. */\n    readonly segmenterNames: readonly string[] | null;\n    /**\n     * Probability above which a segmenter pixel counted as foreground.\n     *\n     * Baked into the graph at fusion time; reported so a caller can say what the\n     * masks mean.\n     */\n    readonly maskThreshold: number;\n    /** Whether the mask gated the crop before the classifier saw it. */\n    readonly maskApplied: boolean;\n}\n\nconst KEY_KIND = \"kind\";\nconst KEY_SDK_VERSION = \"sdk_version\";\nconst KEY_INPUT_SIZE = \"input_size\";\nconst KEY_CROP_SIZE = \"crop_size\";\nconst KEY_CROP_SOURCE = \"crop_source\";\nconst KEY_MAX_DETECTIONS = \"max_detections\";\nconst KEY_CONF_THRESHOLD = \"conf_threshold\";\nconst KEY_IOU_THRESHOLD = \"iou_threshold\";\nconst KEY_APPLY_SOFTMAX = \"apply_softmax\";\nconst KEY_DETECTOR_NAMES = \"detector_names\";\nconst KEY_CLASSIFIER_NAMES = \"classifier_names\";\nconst KEY_SEGMENTER_NAMES = \"segmenter_names\";\nconst KEY_MASK_THRESHOLD = \"mask_threshold\";\nconst KEY_MASK_APPLIED = \"mask_applied\";\n\nconst DYNAMIC = \"dynamic\";\n\n/**\n * Decode a `\"640,640\"` pair.\n *\n * @param raw The encoded pair.\n * @returns `[width, height]`, or `null` when the value is missing or malformed.\n */\nfunction decodeSize(raw: string | undefined): readonly [number, number] | null {\n    if (!raw) return null;\n    const parts = raw.split(\",\");\n    if (parts.length !== 2) return null;\n    const width = Number(parts[0]);\n    const height = Number(parts[1]);\n    if (!Number.isInteger(width) || !Number.isInteger(height)) return null;\n    if (width < 1 || height < 1) return null;\n    return [width, height];\n}\n\n/**\n * Decode a float, falling back when the value is missing or malformed.\n *\n * @param raw The encoded value.\n * @param fallback Value to use when `raw` cannot be read.\n * @returns The parsed number, or `fallback`.\n */\nfunction decodeFloat(raw: string | undefined, fallback: number): number {\n    if (!raw) return fallback;\n    const value = Number(raw);\n    return Number.isFinite(value) ? value : fallback;\n}\n\n/**\n * Read a pipeline spec out of a model's custom metadata.\n *\n * Individual malformed entries fall back to the value a fusion would have used\n * by default — a single bad float is not a reason to reject an otherwise\n * loadable pipeline. A malformed resolution is fatal, because there is no safe\n * default for one.\n *\n * @param metadata A model's custom metadata map, as read by\n *   {@link readModelMetadata}.\n * @returns The decoded spec, or `null` when the model is not a fused pipeline —\n *   it carries no `ovs.kind` entry, or one naming a pipeline kind this version\n *   does not know how to drive.\n */\nexport function readFusionSpec(\n    metadata: Readonly<Record<string, string>> | undefined,\n): FusionSpec | null {\n    if (!metadata) return null;\n\n    const read: Record<string, string> = {};\n    for (const [key, value] of Object.entries(metadata)) {\n        if (key.startsWith(METADATA_PREFIX)) read[key.slice(METADATA_PREFIX.length)] = value;\n    }\n    const kind = read[KEY_KIND];\n    if (kind !== FUSION_KIND_DETECT_CLASSIFY && kind !== FUSION_KIND_DETECT_SEGMENT_CLASSIFY) {\n        return null;\n    }\n\n    const inputSize = decodeSize(read[KEY_INPUT_SIZE]);\n    const cropSize = decodeSize(read[KEY_CROP_SIZE]);\n    if (inputSize === null || cropSize === null) return null;\n\n    const rawMax = read[KEY_MAX_DETECTIONS] ?? DYNAMIC;\n    const parsedMax = Number(rawMax);\n    const maxDetections =\n        rawMax === DYNAMIC || !Number.isInteger(parsedMax) || parsedMax < 1 ? null : parsedMax;\n\n    const cropSource: CropSource =\n        read[KEY_CROP_SOURCE] === \"original\" ? \"original\" : \"detector_input\";\n\n    return {\n        kind,\n        inputSize,\n        cropSize,\n        cropSource,\n        maxDetections,\n        confThreshold: decodeFloat(read[KEY_CONF_THRESHOLD], 0.25),\n        iouThreshold: decodeFloat(read[KEY_IOU_THRESHOLD], 0.45),\n        applySoftmax: (read[KEY_APPLY_SOFTMAX] ?? \"1\") !== \"0\",\n        detectorNames: parseNames(read[KEY_DETECTOR_NAMES]),\n        classifierNames: parseNames(read[KEY_CLASSIFIER_NAMES]),\n        sdkVersion: read[KEY_SDK_VERSION] ?? \"\",\n        needsSourceImage: cropSource === \"original\",\n        hasMasks: kind === FUSION_KIND_DETECT_SEGMENT_CLASSIFY,\n        segmenterNames: parseNames(read[KEY_SEGMENTER_NAMES]),\n        maskThreshold: decodeFloat(read[KEY_MASK_THRESHOLD], 0.5),\n        maskApplied: (read[KEY_MASK_APPLIED] ?? \"1\") !== \"0\",\n    };\n}\n"],"mappings":";;AAoCA,IAAa,IAA8B,mBAW9B,IAAsC,2BAStC,IAAkB,QAGlB,IAAc,UAGd,IAAe,gBAGf,IAAc,mBAGd,IAAY,iBAUZ,IAAe,SAGf,IAAgB,UAGhB,IAAiB,WAGjB,IAAwB,kBAWxB,IAAe,SAEf,IAAe,SA+CtB,IAAW,QACX,IAAkB,eAClB,IAAiB,cACjB,IAAgB,aAChB,IAAkB,eAClB,IAAqB,kBACrB,IAAqB,kBACrB,IAAoB,iBACpB,IAAoB,iBACpB,IAAqB,kBACrB,IAAuB,oBACvB,IAAsB,mBACtB,IAAqB,kBACrB,IAAmB,gBAEnB,IAAU;AAQhB,SAAS,EAAW,GAA2D;CAC3E,IAAI,CAAC,GAAK,OAAO;CACjB,IAAM,IAAQ,EAAI,MAAM,GAAG;CAC3B,IAAI,EAAM,WAAW,GAAG,OAAO;CAC/B,IAAM,IAAQ,OAAO,EAAM,EAAE,GACvB,IAAS,OAAO,EAAM,EAAE;CAG9B,OAFI,CAAC,OAAO,UAAU,CAAK,KAAK,CAAC,OAAO,UAAU,CAAM,KACpD,IAAQ,KAAK,IAAS,IAAU,OAC7B,CAAC,GAAO,CAAM;AACzB;AASA,SAAS,EAAY,GAAyB,GAA0B;CACpE,IAAI,CAAC,GAAK,OAAO;CACjB,IAAM,IAAQ,OAAO,CAAG;CACxB,OAAO,OAAO,SAAS,CAAK,IAAI,IAAQ;AAC5C;AAgBA,SAAgB,EACZ,GACiB;CACjB,IAAI,CAAC,GAAU,OAAO;CAEtB,IAAM,IAA+B,CAAC;CACtC,KAAK,IAAM,CAAC,GAAK,MAAU,OAAO,QAAQ,CAAQ,GAC9C,AAAI,EAAI,WAAA,MAA0B,MAAG,EAAK,EAAI,MAAM,CAAsB,KAAK;CAEnF,IAAM,IAAO,EAAK;CAClB,IAAI,MAAA,qBAAwC,MAAA,2BACxC,OAAO;CAGX,IAAM,IAAY,EAAW,EAAK,EAAe,GAC3C,IAAW,EAAW,EAAK,EAAc;CAC/C,IAAI,MAAc,QAAQ,MAAa,MAAM,OAAO;CAEpD,IAAM,IAAS,EAAK,MAAuB,GACrC,IAAY,OAAO,CAAM,GACzB,IACF,MAAW,KAAW,CAAC,OAAO,UAAU,CAAS,KAAK,IAAY,IAAI,OAAO,GAE3E,IACF,EAAK,OAAqB,aAAa,aAAa;CAExD,OAAO;EACH;EACA;EACA;EACA;EACA;EACA,eAAe,EAAY,EAAK,IAAqB,GAAI;EACzD,cAAc,EAAY,EAAK,IAAoB,GAAI;EACvD,eAAe,EAAK,MAAsB,SAAS;EACnD,eAAe,EAAW,EAAK,EAAmB;EAClD,iBAAiB,EAAW,EAAK,EAAqB;EACtD,YAAY,EAAK,MAAoB;EACrC,kBAAkB,MAAe;EACjC,UAAU,MAAS;EACnB,gBAAgB,EAAW,EAAK,EAAoB;EACpD,eAAe,EAAY,EAAK,IAAqB,EAAG;EACxD,cAAc,EAAK,MAAqB,SAAS;CACrD;AACJ"}

package/dist/vision/index.cjs

@@ -1,2 +1,2 @@
-const e=require("./core/exceptions.cjs"),t=require("./types.cjs"),n=require("./core/timing.cjs"),r=require("./results.cjs"),i=require("./labels.cjs"),a=require("./core/graph.cjs"),o=require("./core/metadata.cjs"),s=require("./core/providers.cjs"),c=require("./core/session.cjs"),l=require("./normalization.cjs"),u=require("./io/image.cjs"),d=require("./fusion.cjs"),f=require("./preprocess/image.cjs"),p=require("./preprocess/pipeline.cjs"),m=require("./postprocess/classification.cjs"),h=require("./postprocess/detection.cjs"),g=require("./postprocess/segmentation.cjs"),_=require("./tasks/base.cjs"),v=require("./tasks/classifier.cjs"),y=require("./tasks/detectClassify.cjs"),b=require("./tasks/detector.cjs"),x=require("./tasks/segmenter.cjs");var S=`0.8.1`;exports.BoundingBox=t.BoundingBox,exports.Boxes=r.Boxes,exports.COCO_CLASSES=i.COCO_CLASSES,exports.CUSTOM_NORMALIZATION=l.CUSTOM_NORMALIZATION,exports.ClassificationResults=r.ClassificationResults,exports.Classifier=v.Classifier,exports.DEFAULT_PROVIDERS=s.DEFAULT_PROVIDERS,exports.DetectClassify=y.DetectClassify,exports.DetectClassifyResults=r.DetectClassifyResults,exports.DetectionResults=r.DetectionResults,exports.Detector=b.Detector,exports.FUSION_KIND_DETECT_CLASSIFY=d.FUSION_KIND_DETECT_CLASSIFY,exports.FusionError=e.FusionError,exports.IDENTITY_MEAN=l.IDENTITY_MEAN,exports.IDENTITY_STD=l.IDENTITY_STD,exports.IMAGENET_MEAN=l.IMAGENET_MEAN,exports.IMAGENET_STD=l.IMAGENET_STD,exports.INPUT_IMAGE=d.INPUT_IMAGE,exports.INPUT_PAD=d.INPUT_PAD,exports.INPUT_SCALE=d.INPUT_SCALE,exports.INPUT_SOURCE=d.INPUT_SOURCE,exports.ImageLoadError=e.ImageLoadError,exports.InferenceError=e.InferenceError,exports.LabelMapError=e.LabelMapError,exports.LetterboxPipeline=p.LetterboxPipeline,exports.METADATA_PREFIX=d.METADATA_PREFIX,exports.Mask=t.Mask,exports.Masks=r.Masks,exports.ModelLoadError=e.ModelLoadError,exports.NoDetectionsError=e.NoDetectionsError,exports.OUTPUT_BOXES=d.OUTPUT_BOXES,exports.OUTPUT_CLASSES=d.OUTPUT_CLASSES,exports.OUTPUT_NUM_DETECTIONS=d.OUTPUT_NUM_DETECTIONS,exports.OUTPUT_PROBS=d.OUTPUT_PROBS,exports.OUTPUT_SCORES=d.OUTPUT_SCORES,exports.OrtSession=c.OrtSession,exports.OrtVisionError=e.OrtVisionError,exports.Probs=r.Probs,exports.ProviderNotAvailableError=e.ProviderNotAvailableError,exports.RGBImage=t.RGBImage,exports.ResizePipeline=p.ResizePipeline,exports.SegmentationResults=r.SegmentationResults,exports.Segmenter=x.Segmenter,exports.SpeedTimer=n.SpeedTimer,exports.VERSION=S,exports.VisionTask=_.VisionTask,exports.batchedNms=h.batchedNms,exports.classificationNumClasses=a.classificationNumClasses,exports.declaredShapesFrom=a.declaredShapesFrom,exports.decodeYolo=h.decodeYolo,exports.decodeYoloAnchors=h.decodeYoloAnchors,exports.decodeYoloSeg=g.decodeYoloSeg,exports.defaultLabels=i.defaultLabels,exports.detectProviders=s.detectProviders,exports.detectionNumClasses=a.detectionNumClasses,exports.fromCv2=f.fromCv2,exports.isUltralyticsClassifier=l.isUltralyticsClassifier,exports.letterbox=f.letterbox,exports.letterboxToTensorData=p.letterboxToTensorData,exports.loadImage=u.loadImage,exports.modelNames=o.modelNames,exports.nms=h.nms,exports.normalize=f.normalize,exports.parseNames=o.parseNames,exports.readFusionSpec=d.readFusionSpec,exports.readModelMetadata=o.readModelMetadata,exports.requireDetections=_.requireDetections,exports.resize=f.resize,exports.resizeToTensorData=p.resizeToTensorData,exports.resolveInputSize=a.resolveInputSize,exports.resolveLabels=i.resolveLabels,exports.resolveNormalization=l.resolveNormalization,exports.resolveProviders=s.resolveProviders,exports.softmax=m.softmax,exports.spatialInputSize=a.spatialInputSize,exports.toCHW=f.toCHW,exports.toCv2=f.toCv2,exports.toFloat32=f.toFloat32,exports.toFloat32Tensor=f.toFloat32Tensor,exports.toTensor=f.toTensor,exports.topK=m.topK,exports.writePlanarFloat32=p.writePlanarFloat32,exports.zeroTensorData=p.zeroTensorData;
+const e=require("./core/exceptions.cjs"),t=require("./types.cjs"),n=require("./core/timing.cjs"),r=require("./results.cjs"),i=require("./labels.cjs"),a=require("./core/dtypes.cjs"),o=require("./core/graph.cjs"),s=require("./core/metadata.cjs"),c=require("./core/providers.cjs"),l=require("./core/session.cjs"),u=require("./normalization.cjs"),d=require("./io/image.cjs"),f=require("./fusion.cjs"),p=require("./preprocess/image.cjs"),m=require("./preprocess/pipeline.cjs"),h=require("./postprocess/classification.cjs"),g=require("./postprocess/detection.cjs"),_=require("./postprocess/segmentation.cjs"),v=require("./tasks/base.cjs"),y=require("./tasks/classifier.cjs"),b=require("./tasks/detectClassify.cjs"),x=require("./tasks/detector.cjs"),S=require("./tasks/segmenter.cjs");var C=`0.11.0`;exports.BoundingBox=t.BoundingBox,exports.Boxes=r.Boxes,exports.COCO_CLASSES=i.COCO_CLASSES,exports.CUSTOM_NORMALIZATION=u.CUSTOM_NORMALIZATION,exports.ClassificationResults=r.ClassificationResults,exports.Classifier=y.Classifier,exports.DEFAULT_MODEL_CACHE=l.DEFAULT_MODEL_CACHE,exports.DEFAULT_PROVIDERS=c.DEFAULT_PROVIDERS,exports.DEFAULT_TENSOR_TYPE=a.DEFAULT_TENSOR_TYPE,exports.DetectClassify=b.DetectClassify,exports.DetectClassifyResults=r.DetectClassifyResults,exports.DetectionResults=r.DetectionResults,exports.Detector=x.Detector,exports.FUSION_KIND_DETECT_CLASSIFY=f.FUSION_KIND_DETECT_CLASSIFY,exports.FUSION_KIND_DETECT_SEGMENT_CLASSIFY=f.FUSION_KIND_DETECT_SEGMENT_CLASSIFY,exports.FusionError=e.FusionError,exports.GRAPH_OPTIMIZATION_KEY=l.GRAPH_OPTIMIZATION_KEY,exports.IDENTITY_MEAN=u.IDENTITY_MEAN,exports.IDENTITY_STD=u.IDENTITY_STD,exports.IMAGENET_MEAN=u.IMAGENET_MEAN,exports.IMAGENET_STD=u.IMAGENET_STD,exports.INPUT_IMAGE=f.INPUT_IMAGE,exports.INPUT_PAD=f.INPUT_PAD,exports.INPUT_SCALE=f.INPUT_SCALE,exports.INPUT_SOURCE=f.INPUT_SOURCE,exports.ImageLoadError=e.ImageLoadError,exports.InferenceError=e.InferenceError,exports.LabelMapError=e.LabelMapError,exports.LetterboxPipeline=m.LetterboxPipeline,exports.METADATA_PREFIX=f.METADATA_PREFIX,exports.Mask=t.Mask,exports.Masks=r.Masks,exports.ModelLoadError=e.ModelLoadError,exports.NoDetectionsError=e.NoDetectionsError,exports.OUTPUT_BOXES=f.OUTPUT_BOXES,exports.OUTPUT_CLASSES=f.OUTPUT_CLASSES,exports.OUTPUT_MASKS=f.OUTPUT_MASKS,exports.OUTPUT_NUM_DETECTIONS=f.OUTPUT_NUM_DETECTIONS,exports.OUTPUT_PROBS=f.OUTPUT_PROBS,exports.OUTPUT_SCORES=f.OUTPUT_SCORES,exports.OrtSession=l.OrtSession,exports.OrtVisionError=e.OrtVisionError,exports.Probs=r.Probs,exports.ProviderNotAvailableError=e.ProviderNotAvailableError,exports.QUANTIZATION_KEY=l.QUANTIZATION_KEY,exports.RGBImage=t.RGBImage,exports.ResizePipeline=m.ResizePipeline,exports.SegmentationResults=r.SegmentationResults,exports.Segmenter=S.Segmenter,exports.SpeedTimer=n.SpeedTimer,exports.VERSION=C,exports.VisionTask=v.VisionTask,exports.asFloat32Array=a.asFloat32Array,exports.batchedNms=g.batchedNms,exports.classificationNumClasses=o.classificationNumClasses,exports.declaredShapesFrom=o.declaredShapesFrom,exports.decodeYolo=g.decodeYolo,exports.decodeYoloAnchors=g.decodeYoloAnchors,exports.decodeYoloSeg=_.decodeYoloSeg,exports.defaultLabels=i.defaultLabels,exports.detectProviders=c.detectProviders,exports.detectionNumClasses=o.detectionNumClasses,exports.fromCv2=p.fromCv2,exports.hasFloat16Array=a.hasFloat16Array,exports.isUltralyticsClassifier=u.isUltralyticsClassifier,exports.letterbox=p.letterbox,exports.letterboxToTensorData=m.letterboxToTensorData,exports.loadImage=d.loadImage,exports.modelNames=s.modelNames,exports.nms=g.nms,exports.normalize=p.normalize,exports.parseNames=s.parseNames,exports.readFusionSpec=f.readFusionSpec,exports.readModelInputTypes=s.readModelInputTypes,exports.readModelMetadata=s.readModelMetadata,exports.requireDetections=v.requireDetections,exports.resize=p.resize,exports.resizeToTensorData=m.resizeToTensorData,exports.resolveInputSize=o.resolveInputSize,exports.resolveLabels=i.resolveLabels,exports.resolveNormalization=u.resolveNormalization,exports.resolveProviders=c.resolveProviders,exports.softmax=h.softmax,exports.spatialInputSize=o.spatialInputSize,exports.tensorTypeFor=a.tensorTypeFor,exports.toCHW=p.toCHW,exports.toCv2=p.toCv2,exports.toFloat32=p.toFloat32,exports.toFloat32Tensor=p.toFloat32Tensor,exports.toTensor=p.toTensor,exports.topK=h.topK,exports.writePlanarFloat32=m.writePlanarFloat32,exports.zeroTensorData=m.zeroTensorData;
 //# sourceMappingURL=index.cjs.map

package/dist/vision/index.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.cjs","names":[],"sources":["../../src/vision/index.ts"],"sourcesContent":["/**\n * `tempest-react-sdk/vision` — browser computer-vision inference with ONNX\n * Runtime Web (classification, detection, segmentation).\n *\n * Vendored from `@mauriciobenjamin700/ort-vision-sdk-web@0.8.1` (MIT, same\n * author) so it ships inside this SDK without an extra package install.\n * `onnxruntime-web` stays an optional peer dependency — install it (and ship\n * the matching `.wasm` files) only when you use this subpath.\n *\n * Do not hand-edit — regenerate with `npm run vendor:vision`.\n */\n\nexport {\n    BoundingBox,\n    Mask,\n    RGBImage,\n    type ClassProbability,\n    type ClassificationResult,\n    type DetectionResult,\n    type SegmentationResult,\n} from \"./types\";\n\nexport {\n    Boxes,\n    ClassificationResults,\n    DetectClassifyResults,\n    DetectionResults,\n    Masks,\n    Probs,\n    SegmentationResults,\n} from \"./results\";\n\nexport {\n    COCO_CLASSES,\n    type LabelSpec,\n    type ResolveLabelsOptions,\n    defaultLabels,\n    resolveLabels,\n} from \"./labels\";\n\nexport {\n    FusionError,\n    ImageLoadError,\n    InferenceError,\n    LabelMapError,\n    ModelLoadError,\n    NoDetectionsError,\n    OrtVisionError,\n    ProviderNotAvailableError,\n} from \"./core/exceptions\";\n\nexport { type ModelSource, type OrtSessionOptions, OrtSession } from \"./core/session\";\nexport {\n    type DeclaredDim,\n    type DeclaredShape,\n    type ResolveInputSizeOptions,\n    classificationNumClasses,\n    declaredShapesFrom,\n    detectionNumClasses,\n    resolveInputSize,\n    spatialInputSize,\n} from \"./core/graph\";\nexport { modelNames, parseNames, readModelMetadata } from \"./core/metadata\";\nexport { DEFAULT_PROVIDERS, detectProviders, resolveProviders } from \"./core/providers\";\nexport { type Speed, SpeedTimer } from \"./core/timing\";\nexport {\n    CUSTOM_NORMALIZATION,\n    IDENTITY_MEAN,\n    IDENTITY_STD,\n    IMAGENET_MEAN,\n    IMAGENET_STD,\n    type Normalization,\n    type ResolvedNormalization,\n    isUltralyticsClassifier,\n    resolveNormalization,\n} from \"./normalization\";\n\nexport { type ImageInput, loadImage } from \"./io/image\";\n\nexport {\n    FUSION_KIND_DETECT_CLASSIFY,\n    INPUT_IMAGE,\n    INPUT_PAD,\n    INPUT_SCALE,\n    INPUT_SOURCE,\n    METADATA_PREFIX,\n    OUTPUT_BOXES,\n    OUTPUT_CLASSES,\n    OUTPUT_NUM_DETECTIONS,\n    OUTPUT_PROBS,\n    OUTPUT_SCORES,\n    type CropSource,\n    type FusionSpec,\n    readFusionSpec,\n} from \"./fusion\";\n\nexport {\n    type LetterboxResult,\n    fromCv2,\n    letterbox,\n    normalize,\n    resize,\n    toCHW,\n    toCv2,\n    toFloat32,\n    toFloat32Tensor,\n    toTensor,\n} from \"./preprocess/image\";\n\nexport {\n    type FusedLetterboxResult,\n    type FusedResizeResult,\n    LetterboxPipeline,\n    ResizePipeline,\n    letterboxToTensorData,\n    resizeToTensorData,\n    writePlanarFloat32,\n    zeroTensorData,\n} from \"./preprocess/pipeline\";\n\nexport { type TopKResult, softmax, topK } from \"./postprocess/classification\";\n\nexport {\n    type DecodeYoloAnchorsOptions,\n    type DecodeYoloOptions,\n    type DecodedAnchors,\n    type DecodedDetection,\n    batchedNms,\n    decodeYolo,\n    decodeYoloAnchors,\n    nms,\n} from \"./postprocess/detection\";\n\nexport {\n    type DecodeYoloSegOptions,\n    type DecodedSegmentation,\n    decodeYoloSeg,\n} from \"./postprocess/segmentation\";\n\nexport { VisionTask, requireDetections } from \"./tasks/base\";\nexport {\n    type ClassifierOptions,\n    type ClassifierPredictOptions,\n    Classifier,\n} from \"./tasks/classifier\";\nexport {\n    type DetectClassifyOptions,\n    type DetectClassifyPredictOptions,\n    DetectClassify,\n} from \"./tasks/detectClassify\";\nexport {\n    type DetectorHead,\n    type DetectorOptions,\n    type DetectorPredictOptions,\n    Detector,\n} from \"./tasks/detector\";\nexport {\n    type SegmenterHead,\n    type SegmenterOptions,\n    type SegmenterPredictOptions,\n    Segmenter,\n} from \"./tasks/segmenter\";\n\nexport const VERSION: string = \"0.8.1\";\n"],"mappings":"2uBAmKA,IAAa,EAAkB"}
+{"version":3,"file":"index.cjs","names":[],"sources":["../../src/vision/index.ts"],"sourcesContent":["/**\n * `tempest-react-sdk/vision` — browser computer-vision inference with ONNX\n * Runtime Web (classification, detection, segmentation).\n *\n * Vendored from `@mauriciobenjamin700/ort-vision-sdk-web@0.11.0` (MIT, same\n * author) so it ships inside this SDK without an extra package install.\n * `onnxruntime-web` stays an optional peer dependency — install it (and ship\n * the matching `.wasm` files) only when you use this subpath.\n *\n * Do not hand-edit — regenerate with `npm run vendor:vision`.\n */\n\nexport {\n    BoundingBox,\n    Mask,\n    RGBImage,\n    type ClassProbability,\n    type ClassificationResult,\n    type DetectionResult,\n    type SegmentationResult,\n} from \"./types\";\n\nexport {\n    Boxes,\n    ClassificationResults,\n    DetectClassifyResults,\n    DetectionResults,\n    Masks,\n    Probs,\n    SegmentationResults,\n} from \"./results\";\n\nexport {\n    COCO_CLASSES,\n    type LabelSpec,\n    type ResolveLabelsOptions,\n    defaultLabels,\n    resolveLabels,\n} from \"./labels\";\n\nexport {\n    FusionError,\n    ImageLoadError,\n    InferenceError,\n    LabelMapError,\n    ModelLoadError,\n    NoDetectionsError,\n    OrtVisionError,\n    ProviderNotAvailableError,\n} from \"./core/exceptions\";\n\nexport {\n    DEFAULT_MODEL_CACHE,\n    GRAPH_OPTIMIZATION_KEY,\n    type ModelSource,\n    type OrtSessionOptions,\n    OrtSession,\n    QUANTIZATION_KEY,\n} from \"./core/session\";\nexport {\n    type DeclaredDim,\n    type DeclaredShape,\n    type ResolveInputSizeOptions,\n    classificationNumClasses,\n    declaredShapesFrom,\n    detectionNumClasses,\n    resolveInputSize,\n    spatialInputSize,\n} from \"./core/graph\";\nexport {\n    DEFAULT_TENSOR_TYPE,\n    type Float16ArrayLike,\n    asFloat32Array,\n    hasFloat16Array,\n    tensorTypeFor,\n} from \"./core/dtypes\";\nexport { modelNames, parseNames, readModelInputTypes, readModelMetadata } from \"./core/metadata\";\nexport {\n    DEFAULT_PROVIDERS,\n    type ProviderSpec,\n    detectProviders,\n    resolveProviders,\n} from \"./core/providers\";\nexport { type Speed, SpeedTimer } from \"./core/timing\";\nexport {\n    CUSTOM_NORMALIZATION,\n    IDENTITY_MEAN,\n    IDENTITY_STD,\n    IMAGENET_MEAN,\n    IMAGENET_STD,\n    type Normalization,\n    type ResolvedNormalization,\n    isUltralyticsClassifier,\n    resolveNormalization,\n} from \"./normalization\";\n\nexport { type ImageInput, loadImage } from \"./io/image\";\n\nexport {\n    FUSION_KIND_DETECT_CLASSIFY,\n    FUSION_KIND_DETECT_SEGMENT_CLASSIFY,\n    INPUT_IMAGE,\n    INPUT_PAD,\n    INPUT_SCALE,\n    INPUT_SOURCE,\n    METADATA_PREFIX,\n    OUTPUT_BOXES,\n    OUTPUT_CLASSES,\n    OUTPUT_MASKS,\n    OUTPUT_NUM_DETECTIONS,\n    OUTPUT_PROBS,\n    OUTPUT_SCORES,\n    type CropSource,\n    type FusionSpec,\n    readFusionSpec,\n} from \"./fusion\";\n\nexport {\n    type LetterboxResult,\n    fromCv2,\n    letterbox,\n    normalize,\n    resize,\n    toCHW,\n    toCv2,\n    toFloat32,\n    toFloat32Tensor,\n    toTensor,\n} from \"./preprocess/image\";\n\nexport {\n    type FusedLetterboxResult,\n    type FusedResizeResult,\n    LetterboxPipeline,\n    ResizePipeline,\n    letterboxToTensorData,\n    resizeToTensorData,\n    writePlanarFloat32,\n    zeroTensorData,\n} from \"./preprocess/pipeline\";\n\nexport { type TopKResult, softmax, topK } from \"./postprocess/classification\";\n\nexport {\n    type DecodeYoloAnchorsOptions,\n    type DecodeYoloOptions,\n    type DecodedAnchors,\n    type DecodedDetection,\n    batchedNms,\n    decodeYolo,\n    decodeYoloAnchors,\n    nms,\n} from \"./postprocess/detection\";\n\nexport {\n    type DecodeYoloSegOptions,\n    type DecodedSegmentation,\n    decodeYoloSeg,\n} from \"./postprocess/segmentation\";\n\nexport { VisionTask, requireDetections } from \"./tasks/base\";\nexport {\n    type ClassifierOptions,\n    type ClassifierPredictOptions,\n    Classifier,\n} from \"./tasks/classifier\";\nexport {\n    type DetectClassifyOptions,\n    type DetectClassifyPredictOptions,\n    DetectClassify,\n} from \"./tasks/detectClassify\";\nexport {\n    type DetectorHead,\n    type DetectorOptions,\n    type DetectorPredictOptions,\n    Detector,\n} from \"./tasks/detector\";\nexport {\n    type SegmenterHead,\n    type SegmenterOptions,\n    type SegmenterPredictOptions,\n    Segmenter,\n} from \"./tasks/segmenter\";\n\nexport const VERSION: string = \"0.11.0\";\n"],"mappings":"0wBAwLA,IAAa,EAAkB"}

package/dist/vision/index.js

@@ -3,26 +3,27 @@ import { BoundingBox as c, Mask as l, RGBImage as u } from "./types.js";
 import { SpeedTimer as d } from "./core/timing.js";
 import { Boxes as f, ClassificationResults as p, DetectClassifyResults as m, DetectionResults as h, Masks as g, Probs as _, SegmentationResults as v } from "./results.js";
 import { COCO_CLASSES as y, defaultLabels as b, resolveLabels as x } from "./labels.js";
-import { classificationNumClasses as S, declaredShapesFrom as C, detectionNumClasses as w, resolveInputSize as T, spatialInputSize as E } from "./core/graph.js";
-import { modelNames as D, parseNames as O, readModelMetadata as k } from "./core/metadata.js";
-import { DEFAULT_PROVIDERS as A, detectProviders as j, resolveProviders as M } from "./core/providers.js";
-import { OrtSession as N } from "./core/session.js";
-import { CUSTOM_NORMALIZATION as P, IDENTITY_MEAN as F, IDENTITY_STD as I, IMAGENET_MEAN as L, IMAGENET_STD as R, isUltralyticsClassifier as z, resolveNormalization as B } from "./normalization.js";
-import { loadImage as V } from "./io/image.js";
-import { FUSION_KIND_DETECT_CLASSIFY as H, INPUT_IMAGE as U, INPUT_PAD as W, INPUT_SCALE as G, INPUT_SOURCE as K, METADATA_PREFIX as q, OUTPUT_BOXES as J, OUTPUT_CLASSES as Y, OUTPUT_NUM_DETECTIONS as X, OUTPUT_PROBS as Z, OUTPUT_SCORES as Q, readFusionSpec as $ } from "./fusion.js";
-import { fromCv2 as ee, letterbox as te, normalize as ne, resize as re, toCHW as ie, toCv2 as ae, toFloat32 as oe, toFloat32Tensor as se, toTensor as ce } from "./preprocess/image.js";
-import { LetterboxPipeline as le, ResizePipeline as ue, letterboxToTensorData as de, resizeToTensorData as fe, writePlanarFloat32 as pe, zeroTensorData as me } from "./preprocess/pipeline.js";
-import { softmax as he, topK as ge } from "./postprocess/classification.js";
-import { batchedNms as _e, decodeYolo as ve, decodeYoloAnchors as ye, nms as be } from "./postprocess/detection.js";
-import { decodeYoloSeg as xe } from "./postprocess/segmentation.js";
-import { VisionTask as Se, requireDetections as Ce } from "./tasks/base.js";
-import { Classifier as we } from "./tasks/classifier.js";
-import { DetectClassify as Te } from "./tasks/detectClassify.js";
-import { Detector as Ee } from "./tasks/detector.js";
-import { Segmenter as De } from "./tasks/segmenter.js";
+import { DEFAULT_TENSOR_TYPE as S, asFloat32Array as C, hasFloat16Array as w, tensorTypeFor as T } from "./core/dtypes.js";
+import { classificationNumClasses as E, declaredShapesFrom as D, detectionNumClasses as O, resolveInputSize as k, spatialInputSize as A } from "./core/graph.js";
+import { modelNames as j, parseNames as M, readModelInputTypes as N, readModelMetadata as P } from "./core/metadata.js";
+import { DEFAULT_PROVIDERS as F, detectProviders as I, resolveProviders as L } from "./core/providers.js";
+import { DEFAULT_MODEL_CACHE as R, GRAPH_OPTIMIZATION_KEY as z, OrtSession as B, QUANTIZATION_KEY as V } from "./core/session.js";
+import { CUSTOM_NORMALIZATION as H, IDENTITY_MEAN as U, IDENTITY_STD as W, IMAGENET_MEAN as G, IMAGENET_STD as K, isUltralyticsClassifier as q, resolveNormalization as J } from "./normalization.js";
+import { loadImage as Y } from "./io/image.js";
+import { FUSION_KIND_DETECT_CLASSIFY as X, FUSION_KIND_DETECT_SEGMENT_CLASSIFY as Z, INPUT_IMAGE as Q, INPUT_PAD as $, INPUT_SCALE as ee, INPUT_SOURCE as te, METADATA_PREFIX as ne, OUTPUT_BOXES as re, OUTPUT_CLASSES as ie, OUTPUT_MASKS as ae, OUTPUT_NUM_DETECTIONS as oe, OUTPUT_PROBS as se, OUTPUT_SCORES as ce, readFusionSpec as le } from "./fusion.js";
+import { fromCv2 as ue, letterbox as de, normalize as fe, resize as pe, toCHW as me, toCv2 as he, toFloat32 as ge, toFloat32Tensor as _e, toTensor as ve } from "./preprocess/image.js";
+import { LetterboxPipeline as ye, ResizePipeline as be, letterboxToTensorData as xe, resizeToTensorData as Se, writePlanarFloat32 as Ce, zeroTensorData as we } from "./preprocess/pipeline.js";
+import { softmax as Te, topK as Ee } from "./postprocess/classification.js";
+import { batchedNms as De, decodeYolo as Oe, decodeYoloAnchors as ke, nms as Ae } from "./postprocess/detection.js";
+import { decodeYoloSeg as je } from "./postprocess/segmentation.js";
+import { VisionTask as Me, requireDetections as Ne } from "./tasks/base.js";
+import { Classifier as Pe } from "./tasks/classifier.js";
+import { DetectClassify as Fe } from "./tasks/detectClassify.js";
+import { Detector as Ie } from "./tasks/detector.js";
+import { Segmenter as Le } from "./tasks/segmenter.js";
 //#region src/vision/index.ts
-var Oe = "0.8.1";
+var Re = "0.11.0";
 //#endregion
-export { c as BoundingBox, f as Boxes, y as COCO_CLASSES, P as CUSTOM_NORMALIZATION, p as ClassificationResults, we as Classifier, A as DEFAULT_PROVIDERS, Te as DetectClassify, m as DetectClassifyResults, h as DetectionResults, Ee as Detector, H as FUSION_KIND_DETECT_CLASSIFY, e as FusionError, F as IDENTITY_MEAN, I as IDENTITY_STD, L as IMAGENET_MEAN, R as IMAGENET_STD, U as INPUT_IMAGE, W as INPUT_PAD, G as INPUT_SCALE, K as INPUT_SOURCE, t as ImageLoadError, n as InferenceError, r as LabelMapError, le as LetterboxPipeline, q as METADATA_PREFIX, l as Mask, g as Masks, i as ModelLoadError, a as NoDetectionsError, J as OUTPUT_BOXES, Y as OUTPUT_CLASSES, X as OUTPUT_NUM_DETECTIONS, Z as OUTPUT_PROBS, Q as OUTPUT_SCORES, N as OrtSession, o as OrtVisionError, _ as Probs, s as ProviderNotAvailableError, u as RGBImage, ue as ResizePipeline, v as SegmentationResults, De as Segmenter, d as SpeedTimer, Oe as VERSION, Se as VisionTask, _e as batchedNms, S as classificationNumClasses, C as declaredShapesFrom, ve as decodeYolo, ye as decodeYoloAnchors, xe as decodeYoloSeg, b as defaultLabels, j as detectProviders, w as detectionNumClasses, ee as fromCv2, z as isUltralyticsClassifier, te as letterbox, de as letterboxToTensorData, V as loadImage, D as modelNames, be as nms, ne as normalize, O as parseNames, $ as readFusionSpec, k as readModelMetadata, Ce as requireDetections, re as resize, fe as resizeToTensorData, T as resolveInputSize, x as resolveLabels, B as resolveNormalization, M as resolveProviders, he as softmax, E as spatialInputSize, ie as toCHW, ae as toCv2, oe as toFloat32, se as toFloat32Tensor, ce as toTensor, ge as topK, pe as writePlanarFloat32, me as zeroTensorData };
+export { c as BoundingBox, f as Boxes, y as COCO_CLASSES, H as CUSTOM_NORMALIZATION, p as ClassificationResults, Pe as Classifier, R as DEFAULT_MODEL_CACHE, F as DEFAULT_PROVIDERS, S as DEFAULT_TENSOR_TYPE, Fe as DetectClassify, m as DetectClassifyResults, h as DetectionResults, Ie as Detector, X as FUSION_KIND_DETECT_CLASSIFY, Z as FUSION_KIND_DETECT_SEGMENT_CLASSIFY, e as FusionError, z as GRAPH_OPTIMIZATION_KEY, U as IDENTITY_MEAN, W as IDENTITY_STD, G as IMAGENET_MEAN, K as IMAGENET_STD, Q as INPUT_IMAGE, $ as INPUT_PAD, ee as INPUT_SCALE, te as INPUT_SOURCE, t as ImageLoadError, n as InferenceError, r as LabelMapError, ye as LetterboxPipeline, ne as METADATA_PREFIX, l as Mask, g as Masks, i as ModelLoadError, a as NoDetectionsError, re as OUTPUT_BOXES, ie as OUTPUT_CLASSES, ae as OUTPUT_MASKS, oe as OUTPUT_NUM_DETECTIONS, se as OUTPUT_PROBS, ce as OUTPUT_SCORES, B as OrtSession, o as OrtVisionError, _ as Probs, s as ProviderNotAvailableError, V as QUANTIZATION_KEY, u as RGBImage, be as ResizePipeline, v as SegmentationResults, Le as Segmenter, d as SpeedTimer, Re as VERSION, Me as VisionTask, C as asFloat32Array, De as batchedNms, E as classificationNumClasses, D as declaredShapesFrom, Oe as decodeYolo, ke as decodeYoloAnchors, je as decodeYoloSeg, b as defaultLabels, I as detectProviders, O as detectionNumClasses, ue as fromCv2, w as hasFloat16Array, q as isUltralyticsClassifier, de as letterbox, xe as letterboxToTensorData, Y as loadImage, j as modelNames, Ae as nms, fe as normalize, M as parseNames, le as readFusionSpec, N as readModelInputTypes, P as readModelMetadata, Ne as requireDetections, pe as resize, Se as resizeToTensorData, k as resolveInputSize, x as resolveLabels, J as resolveNormalization, L as resolveProviders, Te as softmax, A as spatialInputSize, T as tensorTypeFor, me as toCHW, he as toCv2, ge as toFloat32, _e as toFloat32Tensor, ve as toTensor, Ee as topK, Ce as writePlanarFloat32, we as zeroTensorData };
 
 //# sourceMappingURL=index.js.map

package/dist/vision/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","names":[],"sources":["../../src/vision/index.ts"],"sourcesContent":["/**\n * `tempest-react-sdk/vision` — browser computer-vision inference with ONNX\n * Runtime Web (classification, detection, segmentation).\n *\n * Vendored from `@mauriciobenjamin700/ort-vision-sdk-web@0.8.1` (MIT, same\n * author) so it ships inside this SDK without an extra package install.\n * `onnxruntime-web` stays an optional peer dependency — install it (and ship\n * the matching `.wasm` files) only when you use this subpath.\n *\n * Do not hand-edit — regenerate with `npm run vendor:vision`.\n */\n\nexport {\n    BoundingBox,\n    Mask,\n    RGBImage,\n    type ClassProbability,\n    type ClassificationResult,\n    type DetectionResult,\n    type SegmentationResult,\n} from \"./types\";\n\nexport {\n    Boxes,\n    ClassificationResults,\n    DetectClassifyResults,\n    DetectionResults,\n    Masks,\n    Probs,\n    SegmentationResults,\n} from \"./results\";\n\nexport {\n    COCO_CLASSES,\n    type LabelSpec,\n    type ResolveLabelsOptions,\n    defaultLabels,\n    resolveLabels,\n} from \"./labels\";\n\nexport {\n    FusionError,\n    ImageLoadError,\n    InferenceError,\n    LabelMapError,\n    ModelLoadError,\n    NoDetectionsError,\n    OrtVisionError,\n    ProviderNotAvailableError,\n} from \"./core/exceptions\";\n\nexport { type ModelSource, type OrtSessionOptions, OrtSession } from \"./core/session\";\nexport {\n    type DeclaredDim,\n    type DeclaredShape,\n    type ResolveInputSizeOptions,\n    classificationNumClasses,\n    declaredShapesFrom,\n    detectionNumClasses,\n    resolveInputSize,\n    spatialInputSize,\n} from \"./core/graph\";\nexport { modelNames, parseNames, readModelMetadata } from \"./core/metadata\";\nexport { DEFAULT_PROVIDERS, detectProviders, resolveProviders } from \"./core/providers\";\nexport { type Speed, SpeedTimer } from \"./core/timing\";\nexport {\n    CUSTOM_NORMALIZATION,\n    IDENTITY_MEAN,\n    IDENTITY_STD,\n    IMAGENET_MEAN,\n    IMAGENET_STD,\n    type Normalization,\n    type ResolvedNormalization,\n    isUltralyticsClassifier,\n    resolveNormalization,\n} from \"./normalization\";\n\nexport { type ImageInput, loadImage } from \"./io/image\";\n\nexport {\n    FUSION_KIND_DETECT_CLASSIFY,\n    INPUT_IMAGE,\n    INPUT_PAD,\n    INPUT_SCALE,\n    INPUT_SOURCE,\n    METADATA_PREFIX,\n    OUTPUT_BOXES,\n    OUTPUT_CLASSES,\n    OUTPUT_NUM_DETECTIONS,\n    OUTPUT_PROBS,\n    OUTPUT_SCORES,\n    type CropSource,\n    type FusionSpec,\n    readFusionSpec,\n} from \"./fusion\";\n\nexport {\n    type LetterboxResult,\n    fromCv2,\n    letterbox,\n    normalize,\n    resize,\n    toCHW,\n    toCv2,\n    toFloat32,\n    toFloat32Tensor,\n    toTensor,\n} from \"./preprocess/image\";\n\nexport {\n    type FusedLetterboxResult,\n    type FusedResizeResult,\n    LetterboxPipeline,\n    ResizePipeline,\n    letterboxToTensorData,\n    resizeToTensorData,\n    writePlanarFloat32,\n    zeroTensorData,\n} from \"./preprocess/pipeline\";\n\nexport { type TopKResult, softmax, topK } from \"./postprocess/classification\";\n\nexport {\n    type DecodeYoloAnchorsOptions,\n    type DecodeYoloOptions,\n    type DecodedAnchors,\n    type DecodedDetection,\n    batchedNms,\n    decodeYolo,\n    decodeYoloAnchors,\n    nms,\n} from \"./postprocess/detection\";\n\nexport {\n    type DecodeYoloSegOptions,\n    type DecodedSegmentation,\n    decodeYoloSeg,\n} from \"./postprocess/segmentation\";\n\nexport { VisionTask, requireDetections } from \"./tasks/base\";\nexport {\n    type ClassifierOptions,\n    type ClassifierPredictOptions,\n    Classifier,\n} from \"./tasks/classifier\";\nexport {\n    type DetectClassifyOptions,\n    type DetectClassifyPredictOptions,\n    DetectClassify,\n} from \"./tasks/detectClassify\";\nexport {\n    type DetectorHead,\n    type DetectorOptions,\n    type DetectorPredictOptions,\n    Detector,\n} from \"./tasks/detector\";\nexport {\n    type SegmenterHead,\n    type SegmenterOptions,\n    type SegmenterPredictOptions,\n    Segmenter,\n} from \"./tasks/segmenter\";\n\nexport const VERSION: string = \"0.8.1\";\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;AAmKA,IAAa,KAAkB"}
+{"version":3,"file":"index.js","names":[],"sources":["../../src/vision/index.ts"],"sourcesContent":["/**\n * `tempest-react-sdk/vision` — browser computer-vision inference with ONNX\n * Runtime Web (classification, detection, segmentation).\n *\n * Vendored from `@mauriciobenjamin700/ort-vision-sdk-web@0.11.0` (MIT, same\n * author) so it ships inside this SDK without an extra package install.\n * `onnxruntime-web` stays an optional peer dependency — install it (and ship\n * the matching `.wasm` files) only when you use this subpath.\n *\n * Do not hand-edit — regenerate with `npm run vendor:vision`.\n */\n\nexport {\n    BoundingBox,\n    Mask,\n    RGBImage,\n    type ClassProbability,\n    type ClassificationResult,\n    type DetectionResult,\n    type SegmentationResult,\n} from \"./types\";\n\nexport {\n    Boxes,\n    ClassificationResults,\n    DetectClassifyResults,\n    DetectionResults,\n    Masks,\n    Probs,\n    SegmentationResults,\n} from \"./results\";\n\nexport {\n    COCO_CLASSES,\n    type LabelSpec,\n    type ResolveLabelsOptions,\n    defaultLabels,\n    resolveLabels,\n} from \"./labels\";\n\nexport {\n    FusionError,\n    ImageLoadError,\n    InferenceError,\n    LabelMapError,\n    ModelLoadError,\n    NoDetectionsError,\n    OrtVisionError,\n    ProviderNotAvailableError,\n} from \"./core/exceptions\";\n\nexport {\n    DEFAULT_MODEL_CACHE,\n    GRAPH_OPTIMIZATION_KEY,\n    type ModelSource,\n    type OrtSessionOptions,\n    OrtSession,\n    QUANTIZATION_KEY,\n} from \"./core/session\";\nexport {\n    type DeclaredDim,\n    type DeclaredShape,\n    type ResolveInputSizeOptions,\n    classificationNumClasses,\n    declaredShapesFrom,\n    detectionNumClasses,\n    resolveInputSize,\n    spatialInputSize,\n} from \"./core/graph\";\nexport {\n    DEFAULT_TENSOR_TYPE,\n    type Float16ArrayLike,\n    asFloat32Array,\n    hasFloat16Array,\n    tensorTypeFor,\n} from \"./core/dtypes\";\nexport { modelNames, parseNames, readModelInputTypes, readModelMetadata } from \"./core/metadata\";\nexport {\n    DEFAULT_PROVIDERS,\n    type ProviderSpec,\n    detectProviders,\n    resolveProviders,\n} from \"./core/providers\";\nexport { type Speed, SpeedTimer } from \"./core/timing\";\nexport {\n    CUSTOM_NORMALIZATION,\n    IDENTITY_MEAN,\n    IDENTITY_STD,\n    IMAGENET_MEAN,\n    IMAGENET_STD,\n    type Normalization,\n    type ResolvedNormalization,\n    isUltralyticsClassifier,\n    resolveNormalization,\n} from \"./normalization\";\n\nexport { type ImageInput, loadImage } from \"./io/image\";\n\nexport {\n    FUSION_KIND_DETECT_CLASSIFY,\n    FUSION_KIND_DETECT_SEGMENT_CLASSIFY,\n    INPUT_IMAGE,\n    INPUT_PAD,\n    INPUT_SCALE,\n    INPUT_SOURCE,\n    METADATA_PREFIX,\n    OUTPUT_BOXES,\n    OUTPUT_CLASSES,\n    OUTPUT_MASKS,\n    OUTPUT_NUM_DETECTIONS,\n    OUTPUT_PROBS,\n    OUTPUT_SCORES,\n    type CropSource,\n    type FusionSpec,\n    readFusionSpec,\n} from \"./fusion\";\n\nexport {\n    type LetterboxResult,\n    fromCv2,\n    letterbox,\n    normalize,\n    resize,\n    toCHW,\n    toCv2,\n    toFloat32,\n    toFloat32Tensor,\n    toTensor,\n} from \"./preprocess/image\";\n\nexport {\n    type FusedLetterboxResult,\n    type FusedResizeResult,\n    LetterboxPipeline,\n    ResizePipeline,\n    letterboxToTensorData,\n    resizeToTensorData,\n    writePlanarFloat32,\n    zeroTensorData,\n} from \"./preprocess/pipeline\";\n\nexport { type TopKResult, softmax, topK } from \"./postprocess/classification\";\n\nexport {\n    type DecodeYoloAnchorsOptions,\n    type DecodeYoloOptions,\n    type DecodedAnchors,\n    type DecodedDetection,\n    batchedNms,\n    decodeYolo,\n    decodeYoloAnchors,\n    nms,\n} from \"./postprocess/detection\";\n\nexport {\n    type DecodeYoloSegOptions,\n    type DecodedSegmentation,\n    decodeYoloSeg,\n} from \"./postprocess/segmentation\";\n\nexport { VisionTask, requireDetections } from \"./tasks/base\";\nexport {\n    type ClassifierOptions,\n    type ClassifierPredictOptions,\n    Classifier,\n} from \"./tasks/classifier\";\nexport {\n    type DetectClassifyOptions,\n    type DetectClassifyPredictOptions,\n    DetectClassify,\n} from \"./tasks/detectClassify\";\nexport {\n    type DetectorHead,\n    type DetectorOptions,\n    type DetectorPredictOptions,\n    Detector,\n} from \"./tasks/detector\";\nexport {\n    type SegmenterHead,\n    type SegmenterOptions,\n    type SegmenterPredictOptions,\n    Segmenter,\n} from \"./tasks/segmenter\";\n\nexport const VERSION: string = \"0.11.0\";\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;AAwLA,IAAa,KAAkB"}

package/dist/vision/io/image.cjs

@@ -1,2 +1,2 @@
-const e=require("../core/exceptions.cjs"),t=require("../types.cjs"),n=require("../core/canvas.cjs");async function r(r){if(r instanceof t.RGBImage)return r;if(typeof ImageData<`u`&&r instanceof ImageData)return n.imageDataToRGB(r);if(typeof r==`string`)return i(r);if(typeof Blob<`u`&&r instanceof Blob)return a(r);if(typeof HTMLImageElement<`u`&&r instanceof HTMLImageElement)return await o(r),s(r,r.naturalWidth,r.naturalHeight);if(typeof HTMLCanvasElement<`u`&&r instanceof HTMLCanvasElement||typeof OffscreenCanvas<`u`&&r instanceof OffscreenCanvas||typeof ImageBitmap<`u`&&r instanceof ImageBitmap)return s(r,r.width,r.height);throw new e.ImageLoadError(`Unsupported image source type: ${Object.prototype.toString.call(r)}`)}async function i(t){let n;try{n=await fetch(t)}catch(n){throw new e.ImageLoadError(`Failed to fetch image from ${t}: ${n.message}`,{cause:n})}if(!n.ok)throw new e.ImageLoadError(`Failed to fetch image from ${t}: HTTP ${n.status} ${n.statusText}`);return a(await n.blob())}async function a(t){let n;try{n=await createImageBitmap(t)}catch(t){throw new e.ImageLoadError(`Failed to decode image blob: ${t.message}`,{cause:t})}try{return s(n,n.width,n.height)}finally{n.close()}}function o(t){return t.complete&&t.naturalWidth>0?Promise.resolve():new Promise((n,r)=>{let i=()=>{o(),n()},a=()=>{o(),r(new e.ImageLoadError(`Failed to load HTMLImageElement (load event errored)`))},o=()=>{t.removeEventListener(`load`,i),t.removeEventListener(`error`,a)};t.addEventListener(`load`,i,{once:!0}),t.addEventListener(`error`,a,{once:!0})})}function s(t,r,i){if(r===0||i===0)throw new e.ImageLoadError(`Cannot load image with zero dimension (${r}x${i}).`);let a=n.createCanvas(r,i),o=n.get2DContext(a);o.drawImage(t,0,0);let s=o.getImageData(0,0,r,i);return n.imageDataToRGB(s)}exports.loadImage=r;
+const e=require("../core/exceptions.cjs"),t=require("../types.cjs"),n=require("../core/canvas.cjs");async function r(e){return(await i(e)).image}async function i(r){if(r instanceof t.RGBImage)return{image:r,canvas:null,owned:!1};if(typeof ImageData<`u`&&r instanceof ImageData)return{image:n.imageDataToRGBChecked(r).image,canvas:null,owned:!0};if(typeof r==`string`)return a(r);if(typeof Blob<`u`&&r instanceof Blob)return o(r);if(typeof HTMLImageElement<`u`&&r instanceof HTMLImageElement)return await s(r),u(r,r.naturalWidth,r.naturalHeight);if(typeof HTMLVideoElement<`u`&&r instanceof HTMLVideoElement){if(r.readyState<2)throw new e.ImageLoadError(`HTMLVideoElement has no current frame yet (readyState < HAVE_CURRENT_DATA). Wait for its 'loadeddata' event before predicting on it.`);return u(r,r.videoWidth,r.videoHeight,l(r))}if(typeof HTMLCanvasElement<`u`&&r instanceof HTMLCanvasElement||typeof OffscreenCanvas<`u`&&r instanceof OffscreenCanvas||typeof ImageBitmap<`u`&&r instanceof ImageBitmap)return u(r,r.width,r.height);if(typeof VideoFrame<`u`&&r instanceof VideoFrame)return u(r,r.displayWidth,r.displayHeight,r.format!==null&&c.has(r.format));throw new e.ImageLoadError(`Unsupported image source type: ${Object.prototype.toString.call(r)}`)}async function a(t){let n;try{n=await fetch(t)}catch(n){throw new e.ImageLoadError(`Failed to fetch image from ${t}: ${n.message}`,{cause:n})}if(!n.ok)throw new e.ImageLoadError(`Failed to fetch image from ${t}: HTTP ${n.status} ${n.statusText}`);return o(await n.blob())}async function o(t){let n;try{n=await createImageBitmap(t)}catch(t){throw new e.ImageLoadError(`Failed to decode image blob: ${t.message}`,{cause:t})}try{return u(n,n.width,n.height,t.type===`image/jpeg`)}finally{n.close()}}function s(t){return t.complete&&t.naturalWidth>0?Promise.resolve():new Promise((n,r)=>{let i=()=>{o(),n()},a=()=>{o(),r(new e.ImageLoadError(`Failed to load HTMLImageElement (load event errored)`))},o=()=>{t.removeEventListener(`load`,i),t.removeEventListener(`error`,a)};t.addEventListener(`load`,i,{once:!0}),t.addEventListener(`error`,a,{once:!0})})}var c=new Set([`I420`,`I422`,`I444`,`NV12`,`RGBX`,`BGRX`]);function l(e){return typeof MediaStream<`u`&&e.srcObject instanceof MediaStream}function u(r,i,a,o=!1){if(i===0||a===0)throw new e.ImageLoadError(`Cannot load image with zero dimension (${i}x${a}).`);let s=n.createCanvas(i,a),c=n.get2DContext(s);if(c.drawImage(r,0,0,i,a),o)return{image:t.RGBImage.deferred(i,a,()=>n.imageDataToRGBChecked(c.getImageData(0,0,i,a)).image.data),canvas:s,owned:!0};let{image:l,opaque:u}=n.imageDataToRGBChecked(c.getImageData(0,0,i,a));return{image:l,canvas:u?s:null,owned:!0}}exports.loadImage=r,exports.loadImageSource=i;
 //# sourceMappingURL=image.cjs.map

package/dist/vision/io/image.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"image.cjs","names":[],"sources":["../../../src/vision/io/image.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Image loading utilities — convert any supported source into the canonical\n * {@link RGBImage} format.\n *\n * Browser-only: uses `fetch`, `createImageBitmap`, and a Canvas 2D context to\n * decode arbitrary inputs into a tightly-packed `Uint8Array` of HWC RGB pixels.\n */\n\nimport { createCanvas, get2DContext, imageDataToRGB } from \"../core/canvas\";\nimport { ImageLoadError } from \"../core/exceptions\";\nimport { RGBImage } from \"../types\";\n\n/** Anything {@link loadImage} accepts as an image input. */\nexport type ImageInput =\n    | string\n    | Blob\n    | HTMLImageElement\n    | HTMLCanvasElement\n    | OffscreenCanvas\n    | ImageBitmap\n    | ImageData\n    | RGBImage;\n\n/**\n * Load an image from any supported source into a HWC uint8 RGB array.\n *\n * @throws {@link ImageLoadError} if the source cannot be decoded or has an unsupported shape.\n */\nexport async function loadImage(source: ImageInput): Promise<RGBImage> {\n    if (source instanceof RGBImage) {\n        return source;\n    }\n\n    if (typeof ImageData !== \"undefined\" && source instanceof ImageData) {\n        return imageDataToRGB(source);\n    }\n\n    if (typeof source === \"string\") {\n        return loadFromUrl(source);\n    }\n\n    if (typeof Blob !== \"undefined\" && source instanceof Blob) {\n        return loadFromBlob(source);\n    }\n\n    if (typeof HTMLImageElement !== \"undefined\" && source instanceof HTMLImageElement) {\n        await waitForImageElement(source);\n        return drawableToRGB(source, source.naturalWidth, source.naturalHeight);\n    }\n\n    if (typeof HTMLCanvasElement !== \"undefined\" && source instanceof HTMLCanvasElement) {\n        return drawableToRGB(source, source.width, source.height);\n    }\n\n    if (typeof OffscreenCanvas !== \"undefined\" && source instanceof OffscreenCanvas) {\n        return drawableToRGB(source, source.width, source.height);\n    }\n\n    if (typeof ImageBitmap !== \"undefined\" && source instanceof ImageBitmap) {\n        return drawableToRGB(source, source.width, source.height);\n    }\n\n    throw new ImageLoadError(\n        `Unsupported image source type: ${Object.prototype.toString.call(source)}`,\n    );\n}\n\nasync function loadFromUrl(url: string): Promise<RGBImage> {\n    let response: Response;\n    try {\n        response = await fetch(url);\n    } catch (err) {\n        throw new ImageLoadError(`Failed to fetch image from ${url}: ${(err as Error).message}`, {\n            cause: err,\n        });\n    }\n    if (!response.ok) {\n        throw new ImageLoadError(\n            `Failed to fetch image from ${url}: HTTP ${response.status} ${response.statusText}`,\n        );\n    }\n    const blob = await response.blob();\n    return loadFromBlob(blob);\n}\n\nasync function loadFromBlob(blob: Blob): Promise<RGBImage> {\n    let bitmap: ImageBitmap;\n    try {\n        bitmap = await createImageBitmap(blob);\n    } catch (err) {\n        throw new ImageLoadError(`Failed to decode image blob: ${(err as Error).message}`, {\n            cause: err,\n        });\n    }\n    try {\n        return drawableToRGB(bitmap, bitmap.width, bitmap.height);\n    } finally {\n        bitmap.close();\n    }\n}\n\nfunction waitForImageElement(img: HTMLImageElement): Promise<void> {\n    if (img.complete && img.naturalWidth > 0) {\n        return Promise.resolve();\n    }\n    return new Promise<void>((resolve, reject) => {\n        const onLoad = (): void => {\n            cleanup();\n            resolve();\n        };\n        const onError = (): void => {\n            cleanup();\n            reject(new ImageLoadError(\"Failed to load HTMLImageElement (load event errored)\"));\n        };\n        const cleanup = (): void => {\n            img.removeEventListener(\"load\", onLoad);\n            img.removeEventListener(\"error\", onError);\n        };\n        img.addEventListener(\"load\", onLoad, { once: true });\n        img.addEventListener(\"error\", onError, { once: true });\n    });\n}\n\nfunction drawableToRGB(drawable: CanvasImageSource, width: number, height: number): RGBImage {\n    if (width === 0 || height === 0) {\n        throw new ImageLoadError(`Cannot load image with zero dimension (${width}x${height}).`);\n    }\n    const canvas = createCanvas(width, height);\n    const ctx = get2DContext(canvas);\n    ctx.drawImage(drawable, 0, 0);\n    const imageData = ctx.getImageData(0, 0, width, height);\n    return imageDataToRGB(imageData);\n}\n"],"mappings":"oGA6BA,eAAsB,EAAU,EAAuC,CACnE,GAAI,aAAkB,EAAA,SAClB,OAAO,EAGX,GAAI,OAAO,UAAc,KAAe,aAAkB,UACtD,OAAO,EAAA,eAAe,CAAM,EAGhC,GAAI,OAAO,GAAW,SAClB,OAAO,EAAY,CAAM,EAG7B,GAAI,OAAO,KAAS,KAAe,aAAkB,KACjD,OAAO,EAAa,CAAM,EAG9B,GAAI,OAAO,iBAAqB,KAAe,aAAkB,iBAE7D,OADA,MAAM,EAAoB,CAAM,EACzB,EAAc,EAAQ,EAAO,aAAc,EAAO,aAAa,EAW1E,GARI,OAAO,kBAAsB,KAAe,aAAkB,mBAI9D,OAAO,gBAAoB,KAAe,aAAkB,iBAI5D,OAAO,YAAgB,KAAe,aAAkB,YACxD,OAAO,EAAc,EAAQ,EAAO,MAAO,EAAO,MAAM,EAG5D,MAAM,IAAI,EAAA,eACN,kCAAkC,OAAO,UAAU,SAAS,KAAK,CAAM,GAC3E,CACJ,CAEA,eAAe,EAAY,EAAgC,CACvD,IAAI,EACJ,GAAI,CACA,EAAW,MAAM,MAAM,CAAG,CAC9B,OAAS,EAAK,CACV,MAAM,IAAI,EAAA,eAAe,8BAA8B,EAAI,IAAK,EAAc,UAAW,CACrF,MAAO,CACX,CAAC,CACL,CACA,GAAI,CAAC,EAAS,GACV,MAAM,IAAI,EAAA,eACN,8BAA8B,EAAI,SAAS,EAAS,OAAO,GAAG,EAAS,YAC3E,EAGJ,OAAO,EAAa,MADD,EAAS,KAAK,CACT,CAC5B,CAEA,eAAe,EAAa,EAA+B,CACvD,IAAI,EACJ,GAAI,CACA,EAAS,MAAM,kBAAkB,CAAI,CACzC,OAAS,EAAK,CACV,MAAM,IAAI,EAAA,eAAe,gCAAiC,EAAc,UAAW,CAC/E,MAAO,CACX,CAAC,CACL,CACA,GAAI,CACA,OAAO,EAAc,EAAQ,EAAO,MAAO,EAAO,MAAM,CAC5D,QAAU,CACN,EAAO,MAAM,CACjB,CACJ,CAEA,SAAS,EAAoB,EAAsC,CAI/D,OAHI,EAAI,UAAY,EAAI,aAAe,EAC5B,QAAQ,QAAQ,EAEpB,IAAI,SAAe,EAAS,IAAW,CAC1C,IAAM,MAAqB,CACvB,EAAQ,EACR,EAAQ,CACZ,EACM,MAAsB,CACxB,EAAQ,EACR,EAAO,IAAI,EAAA,eAAe,sDAAsD,CAAC,CACrF,EACM,MAAsB,CACxB,EAAI,oBAAoB,OAAQ,CAAM,EACtC,EAAI,oBAAoB,QAAS,CAAO,CAC5C,EACA,EAAI,iBAAiB,OAAQ,EAAQ,CAAE,KAAM,EAAK,CAAC,EACnD,EAAI,iBAAiB,QAAS,EAAS,CAAE,KAAM,EAAK,CAAC,CACzD,CAAC,CACL,CAEA,SAAS,EAAc,EAA6B,EAAe,EAA0B,CACzF,GAAI,IAAU,GAAK,IAAW,EAC1B,MAAM,IAAI,EAAA,eAAe,0CAA0C,EAAM,GAAG,EAAO,GAAG,EAE1F,IAAM,EAAS,EAAA,aAAa,EAAO,CAAM,EACnC,EAAM,EAAA,aAAa,CAAM,EAC/B,EAAI,UAAU,EAAU,EAAG,CAAC,EAC5B,IAAM,EAAY,EAAI,aAAa,EAAG,EAAG,EAAO,CAAM,EACtD,OAAO,EAAA,eAAe,CAAS,CACnC"}
+{"version":3,"file":"image.cjs","names":[],"sources":["../../../src/vision/io/image.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Image loading utilities — convert any supported source into the canonical\n * {@link RGBImage} format.\n *\n * Browser-only: uses `fetch`, `createImageBitmap`, and a Canvas 2D context to\n * decode arbitrary inputs into a tightly-packed `Uint8Array` of HWC RGB pixels.\n */\n\nimport { createCanvas, get2DContext, imageDataToRGBChecked } from \"../core/canvas\";\nimport { ImageLoadError } from \"../core/exceptions\";\nimport { RGBImage } from \"../types\";\n\n/** Anything {@link loadImage} accepts as an image input. */\nexport type ImageInput =\n    | string\n    | Blob\n    | HTMLImageElement\n    | HTMLVideoElement\n    | HTMLCanvasElement\n    | OffscreenCanvas\n    | ImageBitmap\n    | VideoFrame\n    | ImageData\n    | RGBImage;\n\n/**\n * A decoded input, plus what the tasks can reuse from decoding it.\n *\n * Internal to the SDK: the tasks need more than {@link loadImage} returns, and\n * the extra fields are about avoiding work, not about the image itself.\n */\nexport interface LoadedImage {\n    /** The input as the canonical HWC RGB image. */\n    readonly image: RGBImage;\n    /**\n     * A full-resolution canvas holding exactly {@link image}'s pixels, or `null`.\n     *\n     * Decoding any drawable already paints it onto a canvas to read it back.\n     * Handing that canvas to a preprocessing pipeline lets it `drawImage` the\n     * frame straight into the model input, instead of rebuilding RGBA from\n     * {@link image} and `putImageData`-ing it onto a canvas of its own — two\n     * full-resolution passes and an allocation the size of the frame. Only set\n     * when every pixel was opaque: a translucent pixel composites differently\n     * from the RGB that dropping its alpha kept.\n     */\n    readonly canvas: CanvasImageSource | null;\n    /**\n     * Whether {@link image}'s buffer was allocated here rather than passed in.\n     *\n     * An `RGBImage` the caller handed over is theirs, and a video loop may well\n     * refill the same buffer every frame. Anything derived from it lazily would\n     * then read a later frame, so lazy crops are only taken on owned buffers.\n     */\n    readonly owned: boolean;\n}\n\n/**\n * Load an image from any supported source into a HWC uint8 RGB array.\n *\n * A video element must have its current frame available (`readyState >= 2`);\n * the frame showing at the time of the call is the one decoded. A `VideoFrame`\n * is read but not closed — it belongs to the caller.\n *\n * @throws {@link ImageLoadError} if the source cannot be decoded or has an unsupported shape.\n */\nexport async function loadImage(source: ImageInput): Promise<RGBImage> {\n    return (await loadImageSource(source)).image;\n}\n\n/**\n * Decode an input like {@link loadImage}, keeping what preprocessing can reuse.\n *\n * @param source Any {@link ImageInput}.\n * @returns The image, the canvas it was painted on (when reusable), and\n *   whether the SDK owns the pixel buffer.\n * @throws {@link ImageLoadError} if the source cannot be decoded.\n */\nexport async function loadImageSource(source: ImageInput): Promise<LoadedImage> {\n    if (source instanceof RGBImage) {\n        return { image: source, canvas: null, owned: false };\n    }\n\n    if (typeof ImageData !== \"undefined\" && source instanceof ImageData) {\n        return { image: imageDataToRGBChecked(source).image, canvas: null, owned: true };\n    }\n\n    if (typeof source === \"string\") {\n        return loadFromUrl(source);\n    }\n\n    if (typeof Blob !== \"undefined\" && source instanceof Blob) {\n        return loadFromBlob(source);\n    }\n\n    if (typeof HTMLImageElement !== \"undefined\" && source instanceof HTMLImageElement) {\n        await waitForImageElement(source);\n        return drawableToRGB(source, source.naturalWidth, source.naturalHeight);\n    }\n\n    if (typeof HTMLVideoElement !== \"undefined\" && source instanceof HTMLVideoElement) {\n        if (source.readyState < 2) {\n            throw new ImageLoadError(\n                \"HTMLVideoElement has no current frame yet (readyState < HAVE_CURRENT_DATA). \" +\n                    \"Wait for its 'loadeddata' event before predicting on it.\",\n            );\n        }\n        return drawableToRGB(source, source.videoWidth, source.videoHeight, isLiveStream(source));\n    }\n\n    if (typeof HTMLCanvasElement !== \"undefined\" && source instanceof HTMLCanvasElement) {\n        return drawableToRGB(source, source.width, source.height);\n    }\n\n    if (typeof OffscreenCanvas !== \"undefined\" && source instanceof OffscreenCanvas) {\n        return drawableToRGB(source, source.width, source.height);\n    }\n\n    if (typeof ImageBitmap !== \"undefined\" && source instanceof ImageBitmap) {\n        return drawableToRGB(source, source.width, source.height);\n    }\n\n    if (typeof VideoFrame !== \"undefined\" && source instanceof VideoFrame) {\n        return drawableToRGB(\n            source,\n            source.displayWidth,\n            source.displayHeight,\n            source.format !== null && OPAQUE_FRAME_FORMATS.has(source.format),\n        );\n    }\n\n    throw new ImageLoadError(\n        `Unsupported image source type: ${Object.prototype.toString.call(source)}`,\n    );\n}\n\nasync function loadFromUrl(url: string): Promise<LoadedImage> {\n    let response: Response;\n    try {\n        response = await fetch(url);\n    } catch (err) {\n        throw new ImageLoadError(`Failed to fetch image from ${url}: ${(err as Error).message}`, {\n            cause: err,\n        });\n    }\n    if (!response.ok) {\n        throw new ImageLoadError(\n            `Failed to fetch image from ${url}: HTTP ${response.status} ${response.statusText}`,\n        );\n    }\n    const blob = await response.blob();\n    return loadFromBlob(blob);\n}\n\nasync function loadFromBlob(blob: Blob): Promise<LoadedImage> {\n    let bitmap: ImageBitmap;\n    try {\n        bitmap = await createImageBitmap(blob);\n    } catch (err) {\n        throw new ImageLoadError(`Failed to decode image blob: ${(err as Error).message}`, {\n            cause: err,\n        });\n    }\n    try {\n        return drawableToRGB(bitmap, bitmap.width, bitmap.height, blob.type === \"image/jpeg\");\n    } finally {\n        bitmap.close();\n    }\n}\n\nfunction waitForImageElement(img: HTMLImageElement): Promise<void> {\n    if (img.complete && img.naturalWidth > 0) {\n        return Promise.resolve();\n    }\n    return new Promise<void>((resolve, reject) => {\n        const onLoad = (): void => {\n            cleanup();\n            resolve();\n        };\n        const onError = (): void => {\n            cleanup();\n            reject(new ImageLoadError(\"Failed to load HTMLImageElement (load event errored)\"));\n        };\n        const cleanup = (): void => {\n            img.removeEventListener(\"load\", onLoad);\n            img.removeEventListener(\"error\", onError);\n        };\n        img.addEventListener(\"load\", onLoad, { once: true });\n        img.addEventListener(\"error\", onError, { once: true });\n    });\n}\n\n/**\n * `VideoFrame` pixel formats that carry no alpha channel.\n *\n * The `A`-suffixed planar formats and `RGBA`/`BGRA` do; a frame in one of\n * these never has a translucent pixel.\n */\nconst OPAQUE_FRAME_FORMATS: ReadonlySet<string> = new Set([\n    \"I420\",\n    \"I422\",\n    \"I444\",\n    \"NV12\",\n    \"RGBX\",\n    \"BGRX\",\n]);\n\n/**\n * Whether a video element is playing a live capture (camera, screen).\n *\n * `MediaStream` video tracks carry no alpha, so every frame is opaque. A video\n * playing a file may not be — VP9 with alpha exists — so it gets no such\n * promise.\n *\n * @param video The element.\n */\nfunction isLiveStream(video: HTMLVideoElement): boolean {\n    return typeof MediaStream !== \"undefined\" && video.srcObject instanceof MediaStream;\n}\n\n/**\n * Paint a drawable onto a fresh canvas and read it back as RGB.\n *\n * When the source is known to be opaque without looking — a JPEG, a camera\n * frame, an alpha-less `VideoFrame` — the read-back is deferred: the image's\n * pixels come out of the canvas the first time something reads `data`. The\n * canvas is this call's own, so it still holds the frame of this moment\n * however late that read happens. Preprocessing draws from the canvas and\n * never reads `data`, so a loop that only consumes boxes skips the\n * full-resolution `getImageData` altogether.\n *\n * Any other source is read back now, because whether its canvas may stand in\n * for the RGB image depends on every pixel being opaque, and only the read\n * tells.\n *\n * @param drawable The source to decode.\n * @param width Its intrinsic width.\n * @param height Its intrinsic height.\n * @param knownOpaque Whether the source cannot contain a translucent pixel.\n * @returns The decoded image, with the canvas attached when it is opaque.\n * @throws {@link ImageLoadError} for a zero-sized source.\n */\nfunction drawableToRGB(\n    drawable: CanvasImageSource,\n    width: number,\n    height: number,\n    knownOpaque: boolean = false,\n): LoadedImage {\n    if (width === 0 || height === 0) {\n        throw new ImageLoadError(`Cannot load image with zero dimension (${width}x${height}).`);\n    }\n    const canvas = createCanvas(width, height);\n    const ctx = get2DContext(canvas);\n    ctx.drawImage(drawable, 0, 0, width, height);\n    if (knownOpaque) {\n        const image = RGBImage.deferred(\n            width,\n            height,\n            () => imageDataToRGBChecked(ctx.getImageData(0, 0, width, height)).image.data,\n        );\n        return { image, canvas, owned: true };\n    }\n    const { image, opaque } = imageDataToRGBChecked(ctx.getImageData(0, 0, width, height));\n    return { image, canvas: opaque ? canvas : null, owned: true };\n}\n"],"mappings":"oGAkEA,eAAsB,EAAU,EAAuC,CACnE,OAAQ,MAAM,EAAgB,CAAM,EAAA,CAAG,KAC3C,CAUA,eAAsB,EAAgB,EAA0C,CAC5E,GAAI,aAAkB,EAAA,SAClB,MAAO,CAAE,MAAO,EAAQ,OAAQ,KAAM,MAAO,EAAM,EAGvD,GAAI,OAAO,UAAc,KAAe,aAAkB,UACtD,MAAO,CAAE,MAAO,EAAA,sBAAsB,CAAM,CAAC,CAAC,MAAO,OAAQ,KAAM,MAAO,EAAK,EAGnF,GAAI,OAAO,GAAW,SAClB,OAAO,EAAY,CAAM,EAG7B,GAAI,OAAO,KAAS,KAAe,aAAkB,KACjD,OAAO,EAAa,CAAM,EAG9B,GAAI,OAAO,iBAAqB,KAAe,aAAkB,iBAE7D,OADA,MAAM,EAAoB,CAAM,EACzB,EAAc,EAAQ,EAAO,aAAc,EAAO,aAAa,EAG1E,GAAI,OAAO,iBAAqB,KAAe,aAAkB,iBAAkB,CAC/E,GAAI,EAAO,WAAa,EACpB,MAAM,IAAI,EAAA,eACN,sIAEJ,EAEJ,OAAO,EAAc,EAAQ,EAAO,WAAY,EAAO,YAAa,EAAa,CAAM,CAAC,CAC5F,CAUA,GARI,OAAO,kBAAsB,KAAe,aAAkB,mBAI9D,OAAO,gBAAoB,KAAe,aAAkB,iBAI5D,OAAO,YAAgB,KAAe,aAAkB,YACxD,OAAO,EAAc,EAAQ,EAAO,MAAO,EAAO,MAAM,EAG5D,GAAI,OAAO,WAAe,KAAe,aAAkB,WACvD,OAAO,EACH,EACA,EAAO,aACP,EAAO,cACP,EAAO,SAAW,MAAQ,EAAqB,IAAI,EAAO,MAAM,CACpE,EAGJ,MAAM,IAAI,EAAA,eACN,kCAAkC,OAAO,UAAU,SAAS,KAAK,CAAM,GAC3E,CACJ,CAEA,eAAe,EAAY,EAAmC,CAC1D,IAAI,EACJ,GAAI,CACA,EAAW,MAAM,MAAM,CAAG,CAC9B,OAAS,EAAK,CACV,MAAM,IAAI,EAAA,eAAe,8BAA8B,EAAI,IAAK,EAAc,UAAW,CACrF,MAAO,CACX,CAAC,CACL,CACA,GAAI,CAAC,EAAS,GACV,MAAM,IAAI,EAAA,eACN,8BAA8B,EAAI,SAAS,EAAS,OAAO,GAAG,EAAS,YAC3E,EAGJ,OAAO,EAAa,MADD,EAAS,KAAK,CACT,CAC5B,CAEA,eAAe,EAAa,EAAkC,CAC1D,IAAI,EACJ,GAAI,CACA,EAAS,MAAM,kBAAkB,CAAI,CACzC,OAAS,EAAK,CACV,MAAM,IAAI,EAAA,eAAe,gCAAiC,EAAc,UAAW,CAC/E,MAAO,CACX,CAAC,CACL,CACA,GAAI,CACA,OAAO,EAAc,EAAQ,EAAO,MAAO,EAAO,OAAQ,EAAK,OAAS,YAAY,CACxF,QAAU,CACN,EAAO,MAAM,CACjB,CACJ,CAEA,SAAS,EAAoB,EAAsC,CAI/D,OAHI,EAAI,UAAY,EAAI,aAAe,EAC5B,QAAQ,QAAQ,EAEpB,IAAI,SAAe,EAAS,IAAW,CAC1C,IAAM,MAAqB,CACvB,EAAQ,EACR,EAAQ,CACZ,EACM,MAAsB,CACxB,EAAQ,EACR,EAAO,IAAI,EAAA,eAAe,sDAAsD,CAAC,CACrF,EACM,MAAsB,CACxB,EAAI,oBAAoB,OAAQ,CAAM,EACtC,EAAI,oBAAoB,QAAS,CAAO,CAC5C,EACA,EAAI,iBAAiB,OAAQ,EAAQ,CAAE,KAAM,EAAK,CAAC,EACnD,EAAI,iBAAiB,QAAS,EAAS,CAAE,KAAM,EAAK,CAAC,CACzD,CAAC,CACL,CAQA,IAAM,EAA4C,IAAI,IAAI,CACtD,OACA,OACA,OACA,OACA,OACA,MACJ,CAAC,EAWD,SAAS,EAAa,EAAkC,CACpD,OAAO,OAAO,YAAgB,KAAe,EAAM,qBAAqB,WAC5E,CAwBA,SAAS,EACL,EACA,EACA,EACA,EAAuB,GACZ,CACX,GAAI,IAAU,GAAK,IAAW,EAC1B,MAAM,IAAI,EAAA,eAAe,0CAA0C,EAAM,GAAG,EAAO,GAAG,EAE1F,IAAM,EAAS,EAAA,aAAa,EAAO,CAAM,EACnC,EAAM,EAAA,aAAa,CAAM,EAE/B,GADA,EAAI,UAAU,EAAU,EAAG,EAAG,EAAO,CAAM,EACvC,EAMA,MAAO,CAAE,MALK,EAAA,SAAS,SACnB,EACA,MACM,EAAA,sBAAsB,EAAI,aAAa,EAAG,EAAG,EAAO,CAAM,CAAC,CAAC,CAAC,MAAM,IAEpE,EAAO,SAAQ,MAAO,EAAK,EAExC,GAAM,CAAE,QAAO,UAAW,EAAA,sBAAsB,EAAI,aAAa,EAAG,EAAG,EAAO,CAAM,CAAC,EACrF,MAAO,CAAE,QAAO,OAAQ,EAAS,EAAS,KAAM,MAAO,EAAK,CAChE"}

package/dist/vision/io/image.js

@@ -1,17 +1,33 @@
 import { ImageLoadError as e } from "../core/exceptions.js";
 import { RGBImage as t } from "../types.js";
-import { createCanvas as n, get2DContext as r, imageDataToRGB as i } from "../core/canvas.js";
+import { createCanvas as n, get2DContext as r, imageDataToRGBChecked as i } from "../core/canvas.js";
 //#region src/vision/io/image.ts
-async function a(n) {
-	if (n instanceof t) return n;
-	if (typeof ImageData < "u" && n instanceof ImageData) return i(n);
-	if (typeof n == "string") return o(n);
-	if (typeof Blob < "u" && n instanceof Blob) return s(n);
-	if (typeof HTMLImageElement < "u" && n instanceof HTMLImageElement) return await c(n), l(n, n.naturalWidth, n.naturalHeight);
-	if (typeof HTMLCanvasElement < "u" && n instanceof HTMLCanvasElement || typeof OffscreenCanvas < "u" && n instanceof OffscreenCanvas || typeof ImageBitmap < "u" && n instanceof ImageBitmap) return l(n, n.width, n.height);
+async function a(e) {
+	return (await o(e)).image;
+}
+async function o(n) {
+	if (n instanceof t) return {
+		image: n,
+		canvas: null,
+		owned: !1
+	};
+	if (typeof ImageData < "u" && n instanceof ImageData) return {
+		image: i(n).image,
+		canvas: null,
+		owned: !0
+	};
+	if (typeof n == "string") return s(n);
+	if (typeof Blob < "u" && n instanceof Blob) return c(n);
+	if (typeof HTMLImageElement < "u" && n instanceof HTMLImageElement) return await l(n), f(n, n.naturalWidth, n.naturalHeight);
+	if (typeof HTMLVideoElement < "u" && n instanceof HTMLVideoElement) {
+		if (n.readyState < 2) throw new e("HTMLVideoElement has no current frame yet (readyState < HAVE_CURRENT_DATA). Wait for its 'loadeddata' event before predicting on it.");
+		return f(n, n.videoWidth, n.videoHeight, d(n));
+	}
+	if (typeof HTMLCanvasElement < "u" && n instanceof HTMLCanvasElement || typeof OffscreenCanvas < "u" && n instanceof OffscreenCanvas || typeof ImageBitmap < "u" && n instanceof ImageBitmap) return f(n, n.width, n.height);
+	if (typeof VideoFrame < "u" && n instanceof VideoFrame) return f(n, n.displayWidth, n.displayHeight, n.format !== null && u.has(n.format));
 	throw new e(`Unsupported image source type: ${Object.prototype.toString.call(n)}`);
 }
-async function o(t) {
+async function s(t) {
 	let n;
 	try {
 		n = await fetch(t);
@@ -19,9 +35,9 @@ async function o(t) {
 		throw new e(`Failed to fetch image from ${t}: ${n.message}`, { cause: n });
 	}
 	if (!n.ok) throw new e(`Failed to fetch image from ${t}: HTTP ${n.status} ${n.statusText}`);
-	return s(await n.blob());
+	return c(await n.blob());
 }
-async function s(t) {
+async function c(t) {
 	let n;
 	try {
 		n = await createImageBitmap(t);
@@ -29,12 +45,12 @@ async function s(t) {
 		throw new e(`Failed to decode image blob: ${t.message}`, { cause: t });
 	}
 	try {
-		return l(n, n.width, n.height);
+		return f(n, n.width, n.height, t.type === "image/jpeg");
 	} finally {
 		n.close();
 	}
 }
-function c(t) {
+function l(t) {
 	return t.complete && t.naturalWidth > 0 ? Promise.resolve() : new Promise((n, r) => {
 		let i = () => {
 			o(), n();
@@ -46,14 +62,33 @@ function c(t) {
 		t.addEventListener("load", i, { once: !0 }), t.addEventListener("error", a, { once: !0 });
 	});
 }
-function l(t, a, o) {
-	if (a === 0 || o === 0) throw new e(`Cannot load image with zero dimension (${a}x${o}).`);
-	let s = n(a, o), c = r(s);
-	c.drawImage(t, 0, 0);
-	let l = c.getImageData(0, 0, a, o);
-	return i(l);
+var u = /* @__PURE__ */ new Set([
+	"I420",
+	"I422",
+	"I444",
+	"NV12",
+	"RGBX",
+	"BGRX"
+]);
+function d(e) {
+	return typeof MediaStream < "u" && e.srcObject instanceof MediaStream;
+}
+function f(a, o, s, c = !1) {
+	if (o === 0 || s === 0) throw new e(`Cannot load image with zero dimension (${o}x${s}).`);
+	let l = n(o, s), u = r(l);
+	if (u.drawImage(a, 0, 0, o, s), c) return {
+		image: t.deferred(o, s, () => i(u.getImageData(0, 0, o, s)).image.data),
+		canvas: l,
+		owned: !0
+	};
+	let { image: d, opaque: f } = i(u.getImageData(0, 0, o, s));
+	return {
+		image: d,
+		canvas: f ? l : null,
+		owned: !0
+	};
 }
 //#endregion
-export { a as loadImage };
+export { a as loadImage, o as loadImageSource };
 
 //# sourceMappingURL=image.js.map

package/dist/vision/io/image.js.map

@@ -1 +1 @@
-{"version":3,"file":"image.js","names":[],"sources":["../../../src/vision/io/image.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Image loading utilities — convert any supported source into the canonical\n * {@link RGBImage} format.\n *\n * Browser-only: uses `fetch`, `createImageBitmap`, and a Canvas 2D context to\n * decode arbitrary inputs into a tightly-packed `Uint8Array` of HWC RGB pixels.\n */\n\nimport { createCanvas, get2DContext, imageDataToRGB } from \"../core/canvas\";\nimport { ImageLoadError } from \"../core/exceptions\";\nimport { RGBImage } from \"../types\";\n\n/** Anything {@link loadImage} accepts as an image input. */\nexport type ImageInput =\n    | string\n    | Blob\n    | HTMLImageElement\n    | HTMLCanvasElement\n    | OffscreenCanvas\n    | ImageBitmap\n    | ImageData\n    | RGBImage;\n\n/**\n * Load an image from any supported source into a HWC uint8 RGB array.\n *\n * @throws {@link ImageLoadError} if the source cannot be decoded or has an unsupported shape.\n */\nexport async function loadImage(source: ImageInput): Promise<RGBImage> {\n    if (source instanceof RGBImage) {\n        return source;\n    }\n\n    if (typeof ImageData !== \"undefined\" && source instanceof ImageData) {\n        return imageDataToRGB(source);\n    }\n\n    if (typeof source === \"string\") {\n        return loadFromUrl(source);\n    }\n\n    if (typeof Blob !== \"undefined\" && source instanceof Blob) {\n        return loadFromBlob(source);\n    }\n\n    if (typeof HTMLImageElement !== \"undefined\" && source instanceof HTMLImageElement) {\n        await waitForImageElement(source);\n        return drawableToRGB(source, source.naturalWidth, source.naturalHeight);\n    }\n\n    if (typeof HTMLCanvasElement !== \"undefined\" && source instanceof HTMLCanvasElement) {\n        return drawableToRGB(source, source.width, source.height);\n    }\n\n    if (typeof OffscreenCanvas !== \"undefined\" && source instanceof OffscreenCanvas) {\n        return drawableToRGB(source, source.width, source.height);\n    }\n\n    if (typeof ImageBitmap !== \"undefined\" && source instanceof ImageBitmap) {\n        return drawableToRGB(source, source.width, source.height);\n    }\n\n    throw new ImageLoadError(\n        `Unsupported image source type: ${Object.prototype.toString.call(source)}`,\n    );\n}\n\nasync function loadFromUrl(url: string): Promise<RGBImage> {\n    let response: Response;\n    try {\n        response = await fetch(url);\n    } catch (err) {\n        throw new ImageLoadError(`Failed to fetch image from ${url}: ${(err as Error).message}`, {\n            cause: err,\n        });\n    }\n    if (!response.ok) {\n        throw new ImageLoadError(\n            `Failed to fetch image from ${url}: HTTP ${response.status} ${response.statusText}`,\n        );\n    }\n    const blob = await response.blob();\n    return loadFromBlob(blob);\n}\n\nasync function loadFromBlob(blob: Blob): Promise<RGBImage> {\n    let bitmap: ImageBitmap;\n    try {\n        bitmap = await createImageBitmap(blob);\n    } catch (err) {\n        throw new ImageLoadError(`Failed to decode image blob: ${(err as Error).message}`, {\n            cause: err,\n        });\n    }\n    try {\n        return drawableToRGB(bitmap, bitmap.width, bitmap.height);\n    } finally {\n        bitmap.close();\n    }\n}\n\nfunction waitForImageElement(img: HTMLImageElement): Promise<void> {\n    if (img.complete && img.naturalWidth > 0) {\n        return Promise.resolve();\n    }\n    return new Promise<void>((resolve, reject) => {\n        const onLoad = (): void => {\n            cleanup();\n            resolve();\n        };\n        const onError = (): void => {\n            cleanup();\n            reject(new ImageLoadError(\"Failed to load HTMLImageElement (load event errored)\"));\n        };\n        const cleanup = (): void => {\n            img.removeEventListener(\"load\", onLoad);\n            img.removeEventListener(\"error\", onError);\n        };\n        img.addEventListener(\"load\", onLoad, { once: true });\n        img.addEventListener(\"error\", onError, { once: true });\n    });\n}\n\nfunction drawableToRGB(drawable: CanvasImageSource, width: number, height: number): RGBImage {\n    if (width === 0 || height === 0) {\n        throw new ImageLoadError(`Cannot load image with zero dimension (${width}x${height}).`);\n    }\n    const canvas = createCanvas(width, height);\n    const ctx = get2DContext(canvas);\n    ctx.drawImage(drawable, 0, 0);\n    const imageData = ctx.getImageData(0, 0, width, height);\n    return imageDataToRGB(imageData);\n}\n"],"mappings":";;;;AA6BA,eAAsB,EAAU,GAAuC;CACnE,IAAI,aAAkB,GAClB,OAAO;CAGX,IAAI,OAAO,YAAc,OAAe,aAAkB,WACtD,OAAO,EAAe,CAAM;CAGhC,IAAI,OAAO,KAAW,UAClB,OAAO,EAAY,CAAM;CAG7B,IAAI,OAAO,OAAS,OAAe,aAAkB,MACjD,OAAO,EAAa,CAAM;CAG9B,IAAI,OAAO,mBAAqB,OAAe,aAAkB,kBAE7D,OADA,MAAM,EAAoB,CAAM,GACzB,EAAc,GAAQ,EAAO,cAAc,EAAO,aAAa;CAW1E,IARI,OAAO,oBAAsB,OAAe,aAAkB,qBAI9D,OAAO,kBAAoB,OAAe,aAAkB,mBAI5D,OAAO,cAAgB,OAAe,aAAkB,aACxD,OAAO,EAAc,GAAQ,EAAO,OAAO,EAAO,MAAM;CAG5D,MAAM,IAAI,EACN,kCAAkC,OAAO,UAAU,SAAS,KAAK,CAAM,GAC3E;AACJ;AAEA,eAAe,EAAY,GAAgC;CACvD,IAAI;CACJ,IAAI;EACA,IAAW,MAAM,MAAM,CAAG;CAC9B,SAAS,GAAK;EACV,MAAM,IAAI,EAAe,8BAA8B,EAAI,IAAK,EAAc,WAAW,EACrF,OAAO,EACX,CAAC;CACL;CACA,IAAI,CAAC,EAAS,IACV,MAAM,IAAI,EACN,8BAA8B,EAAI,SAAS,EAAS,OAAO,GAAG,EAAS,YAC3E;CAGJ,OAAO,EAAa,MADD,EAAS,KAAK,CACT;AAC5B;AAEA,eAAe,EAAa,GAA+B;CACvD,IAAI;CACJ,IAAI;EACA,IAAS,MAAM,kBAAkB,CAAI;CACzC,SAAS,GAAK;EACV,MAAM,IAAI,EAAe,gCAAiC,EAAc,WAAW,EAC/E,OAAO,EACX,CAAC;CACL;CACA,IAAI;EACA,OAAO,EAAc,GAAQ,EAAO,OAAO,EAAO,MAAM;CAC5D,UAAU;EACN,EAAO,MAAM;CACjB;AACJ;AAEA,SAAS,EAAoB,GAAsC;CAI/D,OAHI,EAAI,YAAY,EAAI,eAAe,IAC5B,QAAQ,QAAQ,IAEpB,IAAI,SAAe,GAAS,MAAW;EAC1C,IAAM,UAAqB;GAEvB,AADA,EAAQ,GACR,EAAQ;EACZ,GACM,UAAsB;GAExB,AADA,EAAQ,GACR,EAAO,IAAI,EAAe,sDAAsD,CAAC;EACrF,GACM,UAAsB;GAExB,AADA,EAAI,oBAAoB,QAAQ,CAAM,GACtC,EAAI,oBAAoB,SAAS,CAAO;EAC5C;EAEA,AADA,EAAI,iBAAiB,QAAQ,GAAQ,EAAE,MAAM,GAAK,CAAC,GACnD,EAAI,iBAAiB,SAAS,GAAS,EAAE,MAAM,GAAK,CAAC;CACzD,CAAC;AACL;AAEA,SAAS,EAAc,GAA6B,GAAe,GAA0B;CACzF,IAAI,MAAU,KAAK,MAAW,GAC1B,MAAM,IAAI,EAAe,0CAA0C,EAAM,GAAG,EAAO,GAAG;CAE1F,IAAM,IAAS,EAAa,GAAO,CAAM,GACnC,IAAM,EAAa,CAAM;CAC/B,EAAI,UAAU,GAAU,GAAG,CAAC;CAC5B,IAAM,IAAY,EAAI,aAAa,GAAG,GAAG,GAAO,CAAM;CACtD,OAAO,EAAe,CAAS;AACnC"}
+{"version":3,"file":"image.js","names":[],"sources":["../../../src/vision/io/image.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Image loading utilities — convert any supported source into the canonical\n * {@link RGBImage} format.\n *\n * Browser-only: uses `fetch`, `createImageBitmap`, and a Canvas 2D context to\n * decode arbitrary inputs into a tightly-packed `Uint8Array` of HWC RGB pixels.\n */\n\nimport { createCanvas, get2DContext, imageDataToRGBChecked } from \"../core/canvas\";\nimport { ImageLoadError } from \"../core/exceptions\";\nimport { RGBImage } from \"../types\";\n\n/** Anything {@link loadImage} accepts as an image input. */\nexport type ImageInput =\n    | string\n    | Blob\n    | HTMLImageElement\n    | HTMLVideoElement\n    | HTMLCanvasElement\n    | OffscreenCanvas\n    | ImageBitmap\n    | VideoFrame\n    | ImageData\n    | RGBImage;\n\n/**\n * A decoded input, plus what the tasks can reuse from decoding it.\n *\n * Internal to the SDK: the tasks need more than {@link loadImage} returns, and\n * the extra fields are about avoiding work, not about the image itself.\n */\nexport interface LoadedImage {\n    /** The input as the canonical HWC RGB image. */\n    readonly image: RGBImage;\n    /**\n     * A full-resolution canvas holding exactly {@link image}'s pixels, or `null`.\n     *\n     * Decoding any drawable already paints it onto a canvas to read it back.\n     * Handing that canvas to a preprocessing pipeline lets it `drawImage` the\n     * frame straight into the model input, instead of rebuilding RGBA from\n     * {@link image} and `putImageData`-ing it onto a canvas of its own — two\n     * full-resolution passes and an allocation the size of the frame. Only set\n     * when every pixel was opaque: a translucent pixel composites differently\n     * from the RGB that dropping its alpha kept.\n     */\n    readonly canvas: CanvasImageSource | null;\n    /**\n     * Whether {@link image}'s buffer was allocated here rather than passed in.\n     *\n     * An `RGBImage` the caller handed over is theirs, and a video loop may well\n     * refill the same buffer every frame. Anything derived from it lazily would\n     * then read a later frame, so lazy crops are only taken on owned buffers.\n     */\n    readonly owned: boolean;\n}\n\n/**\n * Load an image from any supported source into a HWC uint8 RGB array.\n *\n * A video element must have its current frame available (`readyState >= 2`);\n * the frame showing at the time of the call is the one decoded. A `VideoFrame`\n * is read but not closed — it belongs to the caller.\n *\n * @throws {@link ImageLoadError} if the source cannot be decoded or has an unsupported shape.\n */\nexport async function loadImage(source: ImageInput): Promise<RGBImage> {\n    return (await loadImageSource(source)).image;\n}\n\n/**\n * Decode an input like {@link loadImage}, keeping what preprocessing can reuse.\n *\n * @param source Any {@link ImageInput}.\n * @returns The image, the canvas it was painted on (when reusable), and\n *   whether the SDK owns the pixel buffer.\n * @throws {@link ImageLoadError} if the source cannot be decoded.\n */\nexport async function loadImageSource(source: ImageInput): Promise<LoadedImage> {\n    if (source instanceof RGBImage) {\n        return { image: source, canvas: null, owned: false };\n    }\n\n    if (typeof ImageData !== \"undefined\" && source instanceof ImageData) {\n        return { image: imageDataToRGBChecked(source).image, canvas: null, owned: true };\n    }\n\n    if (typeof source === \"string\") {\n        return loadFromUrl(source);\n    }\n\n    if (typeof Blob !== \"undefined\" && source instanceof Blob) {\n        return loadFromBlob(source);\n    }\n\n    if (typeof HTMLImageElement !== \"undefined\" && source instanceof HTMLImageElement) {\n        await waitForImageElement(source);\n        return drawableToRGB(source, source.naturalWidth, source.naturalHeight);\n    }\n\n    if (typeof HTMLVideoElement !== \"undefined\" && source instanceof HTMLVideoElement) {\n        if (source.readyState < 2) {\n            throw new ImageLoadError(\n                \"HTMLVideoElement has no current frame yet (readyState < HAVE_CURRENT_DATA). \" +\n                    \"Wait for its 'loadeddata' event before predicting on it.\",\n            );\n        }\n        return drawableToRGB(source, source.videoWidth, source.videoHeight, isLiveStream(source));\n    }\n\n    if (typeof HTMLCanvasElement !== \"undefined\" && source instanceof HTMLCanvasElement) {\n        return drawableToRGB(source, source.width, source.height);\n    }\n\n    if (typeof OffscreenCanvas !== \"undefined\" && source instanceof OffscreenCanvas) {\n        return drawableToRGB(source, source.width, source.height);\n    }\n\n    if (typeof ImageBitmap !== \"undefined\" && source instanceof ImageBitmap) {\n        return drawableToRGB(source, source.width, source.height);\n    }\n\n    if (typeof VideoFrame !== \"undefined\" && source instanceof VideoFrame) {\n        return drawableToRGB(\n            source,\n            source.displayWidth,\n            source.displayHeight,\n            source.format !== null && OPAQUE_FRAME_FORMATS.has(source.format),\n        );\n    }\n\n    throw new ImageLoadError(\n        `Unsupported image source type: ${Object.prototype.toString.call(source)}`,\n    );\n}\n\nasync function loadFromUrl(url: string): Promise<LoadedImage> {\n    let response: Response;\n    try {\n        response = await fetch(url);\n    } catch (err) {\n        throw new ImageLoadError(`Failed to fetch image from ${url}: ${(err as Error).message}`, {\n            cause: err,\n        });\n    }\n    if (!response.ok) {\n        throw new ImageLoadError(\n            `Failed to fetch image from ${url}: HTTP ${response.status} ${response.statusText}`,\n        );\n    }\n    const blob = await response.blob();\n    return loadFromBlob(blob);\n}\n\nasync function loadFromBlob(blob: Blob): Promise<LoadedImage> {\n    let bitmap: ImageBitmap;\n    try {\n        bitmap = await createImageBitmap(blob);\n    } catch (err) {\n        throw new ImageLoadError(`Failed to decode image blob: ${(err as Error).message}`, {\n            cause: err,\n        });\n    }\n    try {\n        return drawableToRGB(bitmap, bitmap.width, bitmap.height, blob.type === \"image/jpeg\");\n    } finally {\n        bitmap.close();\n    }\n}\n\nfunction waitForImageElement(img: HTMLImageElement): Promise<void> {\n    if (img.complete && img.naturalWidth > 0) {\n        return Promise.resolve();\n    }\n    return new Promise<void>((resolve, reject) => {\n        const onLoad = (): void => {\n            cleanup();\n            resolve();\n        };\n        const onError = (): void => {\n            cleanup();\n            reject(new ImageLoadError(\"Failed to load HTMLImageElement (load event errored)\"));\n        };\n        const cleanup = (): void => {\n            img.removeEventListener(\"load\", onLoad);\n            img.removeEventListener(\"error\", onError);\n        };\n        img.addEventListener(\"load\", onLoad, { once: true });\n        img.addEventListener(\"error\", onError, { once: true });\n    });\n}\n\n/**\n * `VideoFrame` pixel formats that carry no alpha channel.\n *\n * The `A`-suffixed planar formats and `RGBA`/`BGRA` do; a frame in one of\n * these never has a translucent pixel.\n */\nconst OPAQUE_FRAME_FORMATS: ReadonlySet<string> = new Set([\n    \"I420\",\n    \"I422\",\n    \"I444\",\n    \"NV12\",\n    \"RGBX\",\n    \"BGRX\",\n]);\n\n/**\n * Whether a video element is playing a live capture (camera, screen).\n *\n * `MediaStream` video tracks carry no alpha, so every frame is opaque. A video\n * playing a file may not be — VP9 with alpha exists — so it gets no such\n * promise.\n *\n * @param video The element.\n */\nfunction isLiveStream(video: HTMLVideoElement): boolean {\n    return typeof MediaStream !== \"undefined\" && video.srcObject instanceof MediaStream;\n}\n\n/**\n * Paint a drawable onto a fresh canvas and read it back as RGB.\n *\n * When the source is known to be opaque without looking — a JPEG, a camera\n * frame, an alpha-less `VideoFrame` — the read-back is deferred: the image's\n * pixels come out of the canvas the first time something reads `data`. The\n * canvas is this call's own, so it still holds the frame of this moment\n * however late that read happens. Preprocessing draws from the canvas and\n * never reads `data`, so a loop that only consumes boxes skips the\n * full-resolution `getImageData` altogether.\n *\n * Any other source is read back now, because whether its canvas may stand in\n * for the RGB image depends on every pixel being opaque, and only the read\n * tells.\n *\n * @param drawable The source to decode.\n * @param width Its intrinsic width.\n * @param height Its intrinsic height.\n * @param knownOpaque Whether the source cannot contain a translucent pixel.\n * @returns The decoded image, with the canvas attached when it is opaque.\n * @throws {@link ImageLoadError} for a zero-sized source.\n */\nfunction drawableToRGB(\n    drawable: CanvasImageSource,\n    width: number,\n    height: number,\n    knownOpaque: boolean = false,\n): LoadedImage {\n    if (width === 0 || height === 0) {\n        throw new ImageLoadError(`Cannot load image with zero dimension (${width}x${height}).`);\n    }\n    const canvas = createCanvas(width, height);\n    const ctx = get2DContext(canvas);\n    ctx.drawImage(drawable, 0, 0, width, height);\n    if (knownOpaque) {\n        const image = RGBImage.deferred(\n            width,\n            height,\n            () => imageDataToRGBChecked(ctx.getImageData(0, 0, width, height)).image.data,\n        );\n        return { image, canvas, owned: true };\n    }\n    const { image, opaque } = imageDataToRGBChecked(ctx.getImageData(0, 0, width, height));\n    return { image, canvas: opaque ? canvas : null, owned: true };\n}\n"],"mappings":";;;;AAkEA,eAAsB,EAAU,GAAuC;CACnE,QAAQ,MAAM,EAAgB,CAAM,EAAA,CAAG;AAC3C;AAUA,eAAsB,EAAgB,GAA0C;CAC5E,IAAI,aAAkB,GAClB,OAAO;EAAE,OAAO;EAAQ,QAAQ;EAAM,OAAO;CAAM;CAGvD,IAAI,OAAO,YAAc,OAAe,aAAkB,WACtD,OAAO;EAAE,OAAO,EAAsB,CAAM,CAAC,CAAC;EAAO,QAAQ;EAAM,OAAO;CAAK;CAGnF,IAAI,OAAO,KAAW,UAClB,OAAO,EAAY,CAAM;CAG7B,IAAI,OAAO,OAAS,OAAe,aAAkB,MACjD,OAAO,EAAa,CAAM;CAG9B,IAAI,OAAO,mBAAqB,OAAe,aAAkB,kBAE7D,OADA,MAAM,EAAoB,CAAM,GACzB,EAAc,GAAQ,EAAO,cAAc,EAAO,aAAa;CAG1E,IAAI,OAAO,mBAAqB,OAAe,aAAkB,kBAAkB;EAC/E,IAAI,EAAO,aAAa,GACpB,MAAM,IAAI,EACN,sIAEJ;EAEJ,OAAO,EAAc,GAAQ,EAAO,YAAY,EAAO,aAAa,EAAa,CAAM,CAAC;CAC5F;CAUA,IARI,OAAO,oBAAsB,OAAe,aAAkB,qBAI9D,OAAO,kBAAoB,OAAe,aAAkB,mBAI5D,OAAO,cAAgB,OAAe,aAAkB,aACxD,OAAO,EAAc,GAAQ,EAAO,OAAO,EAAO,MAAM;CAG5D,IAAI,OAAO,aAAe,OAAe,aAAkB,YACvD,OAAO,EACH,GACA,EAAO,cACP,EAAO,eACP,EAAO,WAAW,QAAQ,EAAqB,IAAI,EAAO,MAAM,CACpE;CAGJ,MAAM,IAAI,EACN,kCAAkC,OAAO,UAAU,SAAS,KAAK,CAAM,GAC3E;AACJ;AAEA,eAAe,EAAY,GAAmC;CAC1D,IAAI;CACJ,IAAI;EACA,IAAW,MAAM,MAAM,CAAG;CAC9B,SAAS,GAAK;EACV,MAAM,IAAI,EAAe,8BAA8B,EAAI,IAAK,EAAc,WAAW,EACrF,OAAO,EACX,CAAC;CACL;CACA,IAAI,CAAC,EAAS,IACV,MAAM,IAAI,EACN,8BAA8B,EAAI,SAAS,EAAS,OAAO,GAAG,EAAS,YAC3E;CAGJ,OAAO,EAAa,MADD,EAAS,KAAK,CACT;AAC5B;AAEA,eAAe,EAAa,GAAkC;CAC1D,IAAI;CACJ,IAAI;EACA,IAAS,MAAM,kBAAkB,CAAI;CACzC,SAAS,GAAK;EACV,MAAM,IAAI,EAAe,gCAAiC,EAAc,WAAW,EAC/E,OAAO,EACX,CAAC;CACL;CACA,IAAI;EACA,OAAO,EAAc,GAAQ,EAAO,OAAO,EAAO,QAAQ,EAAK,SAAS,YAAY;CACxF,UAAU;EACN,EAAO,MAAM;CACjB;AACJ;AAEA,SAAS,EAAoB,GAAsC;CAI/D,OAHI,EAAI,YAAY,EAAI,eAAe,IAC5B,QAAQ,QAAQ,IAEpB,IAAI,SAAe,GAAS,MAAW;EAC1C,IAAM,UAAqB;GAEvB,AADA,EAAQ,GACR,EAAQ;EACZ,GACM,UAAsB;GAExB,AADA,EAAQ,GACR,EAAO,IAAI,EAAe,sDAAsD,CAAC;EACrF,GACM,UAAsB;GAExB,AADA,EAAI,oBAAoB,QAAQ,CAAM,GACtC,EAAI,oBAAoB,SAAS,CAAO;EAC5C;EAEA,AADA,EAAI,iBAAiB,QAAQ,GAAQ,EAAE,MAAM,GAAK,CAAC,GACnD,EAAI,iBAAiB,SAAS,GAAS,EAAE,MAAM,GAAK,CAAC;CACzD,CAAC;AACL;AAQA,IAAM,oBAA4C,IAAI,IAAI;CACtD;CACA;CACA;CACA;CACA;CACA;AACJ,CAAC;AAWD,SAAS,EAAa,GAAkC;CACpD,OAAO,OAAO,cAAgB,OAAe,EAAM,qBAAqB;AAC5E;AAwBA,SAAS,EACL,GACA,GACA,GACA,IAAuB,IACZ;CACX,IAAI,MAAU,KAAK,MAAW,GAC1B,MAAM,IAAI,EAAe,0CAA0C,EAAM,GAAG,EAAO,GAAG;CAE1F,IAAM,IAAS,EAAa,GAAO,CAAM,GACnC,IAAM,EAAa,CAAM;CAE/B,IADA,EAAI,UAAU,GAAU,GAAG,GAAG,GAAO,CAAM,GACvC,GAMA,OAAO;EAAE,OALK,EAAS,SACnB,GACA,SACM,EAAsB,EAAI,aAAa,GAAG,GAAG,GAAO,CAAM,CAAC,CAAC,CAAC,MAAM,IAEpE;EAAO;EAAQ,OAAO;CAAK;CAExC,IAAM,EAAE,UAAO,cAAW,EAAsB,EAAI,aAAa,GAAG,GAAG,GAAO,CAAM,CAAC;CACrF,OAAO;EAAE;EAAO,QAAQ,IAAS,IAAS;EAAM,OAAO;CAAK;AAChE"}

package/dist/vision/postprocess/detection.cjs

@@ -1,2 +1,2 @@
-const e=require("../types.cjs");function t(e,t,n){let r=t.length;if(r===0)return new Int32Array;let i=new Float32Array(r);for(let t=0;t<r;t++){let n=e[t*4],r=e[t*4+1],a=e[t*4+2],o=e[t*4+3];i[t]=Math.max(0,a-n)*Math.max(0,o-r)}let a=Array(r);for(let e=0;e<r;e++)a[e]=e;a.sort((e,n)=>t[n]-t[e]||e-n);let o=new Uint8Array(r),s=[];for(let t=0;t<a.length;t++){let r=a[t];if(o[r])continue;s.push(r);let c=e[r*4],l=e[r*4+1],u=e[r*4+2],d=e[r*4+3],f=i[r];for(let r=t+1;r<a.length;r++){let t=a[r];if(o[t])continue;let s=e[t*4],p=e[t*4+1],m=e[t*4+2],h=e[t*4+3],g=Math.max(c,s),_=Math.max(l,p),v=Math.min(u,m),y=Math.min(d,h),b=Math.max(0,v-g)*Math.max(0,y-_),x=f+i[t]-b;(x>0?b/x:0)>n&&(o[t]=1)}}return Int32Array.from(s)}function n(e,n,r,i){if(n.length===0)return new Int32Array;let a=new Map;for(let e=0;e<r.length;e++){let t=r[e],n=a.get(t);n===void 0?a.set(t,[e]):n.push(e)}let o=[];for(let r of a.values()){let a=r.length,s=new Float32Array(a*4),c=new Float32Array(a);for(let t=0;t<a;t++){let i=r[t];s[t*4]=e[i*4],s[t*4+1]=e[i*4+1],s[t*4+2]=e[i*4+2],s[t*4+3]=e[i*4+3],c[t]=n[i]}let l=t(s,c,i);for(let e=0;e<l.length;e++)o.push(r[l[e]])}return o.sort((e,t)=>n[t]-n[e]||e-t),Int32Array.from(o)}function r(e,t,r){let a=t;if(a.length===3){if(a[0]!==1)throw Error(`decodeYoloAnchors: expected batch size 1, got ${a[0]}.`);a=[a[1],a[2]]}if(a.length!==2)throw Error(`decodeYoloAnchors: expected 2-D output after batch removal, got dims=${JSON.stringify(t)}.`);let o=a[0],s=a[1],{numClasses:c,originalWidth:l,originalHeight:u,padLeft:d,padTop:f,scale:p,confThreshold:m,iouThreshold:h,maxDetections:g}=r;if(c<1||c+4>o)throw Error(`decodeYoloAnchors: invalid numClasses=${c} for channels=${o}.`);if(e.length!==o*s)throw Error(`decodeYoloAnchors: data length ${e.length} does not match channels*numAnchors=${o*s}.`);let _=[];for(let t=0;t<s;t++){let n=0,r=-1/0;for(let i=0;i<c;i++){let a=e[(4+i)*s+t];a!==void 0&&a>r&&(r=a,n=i)}if(r<m)continue;let i=e[t],a=e[s+t],o=e[2*s+t],h=e[3*s+t],g=i-o/2,v=a-h/2,y=i+o/2,b=a+h/2;g=(g-d)/p,v=(v-f)/p,y=(y-d)/p,b=(b-f)/p,g=Math.max(0,Math.min(l,g)),v=Math.max(0,Math.min(u,v)),y=Math.max(0,Math.min(l,y)),b=Math.max(0,Math.min(u,b)),_.push({anchorIdx:t,x1:g,y1:v,x2:y,y2:b,classId:n,confidence:r})}if(_.length===0)return i();let v=new Float32Array(_.length*4),y=new Float32Array(_.length),b=new Int32Array(_.length);for(let e=0;e<_.length;e++){let t=_[e];v[e*4]=t.x1,v[e*4+1]=t.y1,v[e*4+2]=t.x2,v[e*4+3]=t.y2,y[e]=t.confidence,b[e]=t.classId}let x=n(v,y,b,h);if(x.length===0)return i();let S=Array.from(x).slice(0,g),C=S.length,w=new Int32Array(C),T=new Float32Array(C*4),E=new Int32Array(C),D=new Float32Array(C);for(let e=0;e<C;e++){let t=_[S[e]];w[e]=t.anchorIdx,T[e*4]=t.x1,T[e*4+1]=t.y1,T[e*4+2]=t.x2,T[e*4+3]=t.y2,E[e]=t.classId,D[e]=t.confidence}return{anchorIndices:w,boxesXyxy:T,classIds:E,confidences:D}}function i(){return{anchorIndices:new Int32Array,boxesXyxy:new Float32Array,classIds:new Int32Array,confidences:new Float32Array}}function a(t,n,i){let a=n.length===3?n[1]:n[0];if(a===void 0||a<5)throw Error(`decodeYolo: invalid output channel count ${a} (expected >= 5).`);let o=r(t,n,{numClasses:a-4,...i}),s=[];for(let t=0;t<o.classIds.length;t++)s.push({bbox:new e.BoundingBox(o.boxesXyxy[t*4],o.boxesXyxy[t*4+1],o.boxesXyxy[t*4+2],o.boxesXyxy[t*4+3]),classId:o.classIds[t],confidence:o.confidences[t]});return s}exports.batchedNms=n,exports.decodeYolo=a,exports.decodeYoloAnchors=r,exports.nms=t;
+const e=require("../types.cjs");function t(e,t,n){let r=t.length;if(r===0)return new Int32Array;let i=Array(r);for(let e=0;e<r;e++)i[e]=e;i.sort((e,n)=>t[n]-t[e]||e-n);let a=new Float32Array(r*4),o=new Float32Array(r);for(let t=0;t<r;t++){let n=i[t]*4,r=e[n],s=e[n+1],c=e[n+2],l=e[n+3];a[t*4]=r,a[t*4+1]=s,a[t*4+2]=c,a[t*4+3]=l,o[t]=Math.max(0,c-r)*Math.max(0,l-s)}let s=n>=0,c=new Uint8Array(r),l=[];for(let e=0;e<r;e++){if(c[e])continue;l.push(i[e]);let t=a[e*4],u=a[e*4+1],d=a[e*4+2],f=a[e*4+3],p=o[e];for(let i=e+1;i<r;i++){if(c[i])continue;let e=a[i*4],r=a[i*4+1],l=a[i*4+2],m=a[i*4+3],h=Math.max(0,Math.min(d,l)-Math.max(t,e));if(h===0&&s)continue;let g=h*Math.max(0,Math.min(f,m)-Math.max(u,r)),_=p+o[i]-g;(_>0?g/_:0)>n&&(c[i]=1)}}return Int32Array.from(l)}function n(e,n,r,i){if(n.length===0)return new Int32Array;let a=new Map;for(let e=0;e<r.length;e++){let t=r[e],n=a.get(t);n===void 0?a.set(t,[e]):n.push(e)}let o=[];for(let r of a.values()){let a=r.length,s=new Float32Array(a*4),c=new Float32Array(a);for(let t=0;t<a;t++){let i=r[t];s[t*4]=e[i*4],s[t*4+1]=e[i*4+1],s[t*4+2]=e[i*4+2],s[t*4+3]=e[i*4+3],c[t]=n[i]}let l=t(s,c,i);for(let e=0;e<l.length;e++)o.push(r[l[e]])}return o.sort((e,t)=>n[t]-n[e]||e-t),Int32Array.from(o)}function r(e,t,r){let a=t;if(a.length===3){if(a[0]!==1)throw Error(`decodeYoloAnchors: expected batch size 1, got ${a[0]}.`);a=[a[1],a[2]]}if(a.length!==2)throw Error(`decodeYoloAnchors: expected 2-D output after batch removal, got dims=${JSON.stringify(t)}.`);let o=a[0],s=a[1],{numClasses:c,originalWidth:l,originalHeight:u,padLeft:d,padTop:f,scale:p,confThreshold:m,iouThreshold:h,maxDetections:g}=r;if(c<1||c+4>o)throw Error(`decodeYoloAnchors: invalid numClasses=${c} for channels=${o}.`);if(e.length!==o*s)throw Error(`decodeYoloAnchors: data length ${e.length} does not match channels*numAnchors=${o*s}.`);let _=new Uint8Array(s),v=(4+c)*s;for(let t=4*s;t<v;t+=s)for(let n=0;n<s;n++)e[t+n]>=m&&(_[n]=1);let y=0,b=new Int32Array(s);for(let e=0;e<s;e++)_[e]===1&&(b[y++]=e);if(y===0)return i();let x=new Float32Array(y*4),S=new Float32Array(y),C=new Int32Array(y);for(let t=0;t<y;t++){let n=b[t],r=e[n],i=e[s+n],a=e[2*s+n],o=e[3*s+n],m=(r-a/2-d)/p,h=(i-o/2-f)/p,g=(r+a/2-d)/p,_=(i+o/2-f)/p;x[t*4]=Math.max(0,Math.min(l,m)),x[t*4+1]=Math.max(0,Math.min(u,h)),x[t*4+2]=Math.max(0,Math.min(l,g)),x[t*4+3]=Math.max(0,Math.min(u,_));let v=-1/0,y=0;for(let t=0,r=4*s+n;t<c;t++,r+=s){let n=e[r];n>v&&(v=n,y=t)}S[t]=v,C[t]=y}let w=n(x,S,C,h);if(w.length===0)return i();let T=Math.min(w.length,g),E=new Int32Array(T),D=new Float32Array(T*4),O=new Int32Array(T),k=new Float32Array(T);for(let e=0;e<T;e++){let t=w[e];E[e]=b[t],D[e*4]=x[t*4],D[e*4+1]=x[t*4+1],D[e*4+2]=x[t*4+2],D[e*4+3]=x[t*4+3],O[e]=C[t],k[e]=S[t]}return{anchorIndices:E,boxesXyxy:D,classIds:O,confidences:k}}function i(){return{anchorIndices:new Int32Array,boxesXyxy:new Float32Array,classIds:new Int32Array,confidences:new Float32Array}}function a(t,n,i){let a=n.length===3?n[1]:n[0];if(a===void 0||a<5)throw Error(`decodeYolo: invalid output channel count ${a} (expected >= 5).`);let o=r(t,n,{numClasses:a-4,...i}),s=[];for(let t=0;t<o.classIds.length;t++)s.push({bbox:new e.BoundingBox(o.boxesXyxy[t*4],o.boxesXyxy[t*4+1],o.boxesXyxy[t*4+2],o.boxesXyxy[t*4+3]),classId:o.classIds[t],confidence:o.confidences[t]});return s}exports.batchedNms=n,exports.decodeYolo=a,exports.decodeYoloAnchors=r,exports.nms=t;
 //# sourceMappingURL=detection.cjs.map