@mauriciobenjamin700/ort-vision-sdk-web 0.2.1

package/dist/tasks/detector.d.ts

@@ -0,0 +1,102 @@
+/**
+ * Object detection task using anchor-free YOLO ONNX models (v8/v9/v10/v11/v12).
+ */
+import { type ModelSource, type OrtSessionOptions } from "../core/session.js";
+import { type ImageInput } from "../io/image.js";
+import { type LabelSpec } from "../labels.js";
+import { DetectionResults } from "../results.js";
+import { VisionTask } from "./base.js";
+/**
+ * Decoder family for the model's detection head.
+ *
+ * - `"yolo"`: anchor-free YOLO head with output shape `[1, 4 + nc, N]` —
+ *   covers YOLOv8, v9, v10, v11, v12, v26 detect exports.
+ *
+ * The SDK does **not** auto-detect the head from the model — the caller is
+ * responsible for picking a head that matches their export. Future families
+ * (v5/v6/v7 with `[1, N, 5+nc]`) will be added as new literal members.
+ */
+export type DetectorHead = "yolo";
+export interface DetectorOptions extends OrtSessionOptions {
+    /**
+     * Decoder family for the detection head. Default `"yolo"` covers
+     * YOLOv8/v9/v10/v11/v12/v26.
+     */
+    readonly head?: DetectorHead;
+    /** Class label spec — see {@link resolveLabels}. Defaults to the COCO 80-class preset. */
+    readonly labels?: LabelSpec;
+    /** Number of classes — used to validate the supplied labels. */
+    readonly numClasses?: number;
+    /** Model input `[width, height]` for letterboxing. Defaults to `[640, 640]`. */
+    readonly inputSize?: readonly [number, number];
+    /** Default minimum class score to keep a candidate. */
+    readonly confThreshold?: number;
+    /** Default IoU threshold for non-maximum suppression. */
+    readonly iouThreshold?: number;
+    /** Maximum number of detections per image. */
+    readonly maxDetections?: number;
+}
+export interface DetectorPredictOptions {
+    /** Override the default confidence threshold. */
+    readonly confThreshold?: number;
+    /** Override the default IoU threshold. */
+    readonly iouThreshold?: number;
+    /**
+     * If set, keep only detections whose `classId` is in this list.
+     * Mirrors Ultralytics' `model.predict(img, classes=[0, 16])`.
+     */
+    readonly classes?: readonly number[];
+}
+/**
+ * Object detector for anchor-free YOLO ONNX models (v8/v9/v10/v11/v12).
+ *
+ * `predict()` returns `Promise<DetectionResults[]>` (length 1 for a single
+ * image), mirroring Ultralytics' `YOLO("img.jpg")`. Iterate the envelope for
+ * per-instance dataclasses, or use the bulk `boxes` view (`.xyxy`, `.xywh`,
+ * `.xyxyn`, `.xywhn`, `.cls`, `.conf`).
+ *
+ * @example
+ * ```typescript
+ * const det = await Detector.create("/models/yolov8n.onnx");
+ * const results = await det.predict("/images/street.jpg");
+ * const r = results[0];
+ * console.log(r.boxes.xyxy, r.boxes.cls, r.boxes.conf, r.names);
+ * for (const d of r) {
+ *   console.log(d.cls, d.conf, d.box.xyxy);
+ * }
+ * ```
+ */
+export declare class Detector extends VisionTask {
+    private readonly _head;
+    private readonly _labels;
+    private readonly _names;
+    private readonly _inputSize;
+    private readonly _confThreshold;
+    private readonly _iouThreshold;
+    private readonly _maxDetections;
+    private constructor();
+    /** Load the model and resolve labels. */
+    static create(model: ModelSource, options?: DetectorOptions): Promise<Detector>;
+    /** The decoder family used to interpret the model's output. */
+    get head(): DetectorHead;
+    /** Class labels indexed by class id. */
+    get labels(): readonly string[];
+    /** Class id → class name dict (matches Ultralytics' `model.names`). */
+    get names(): Readonly<Record<number, string>>;
+    /** Number of classes the model predicts. */
+    get numClasses(): number;
+    /**
+     * Alias for {@link predict} — call the detector like a torch `nn.Module`.
+     *
+     * Use as `det.call(img)` since JavaScript class instances are not callable;
+     * for direct invocation, prefer `det.predict(img)`. The full
+     * {@link DetectorPredictOptions} (including `classes`) is supported.
+     */
+    call(image: ImageInput, options?: DetectorPredictOptions): Promise<DetectionResults[]>;
+    /** Run detection on a single image. */
+    predict(image: ImageInput, options?: DetectorPredictOptions): Promise<DetectionResults[]>;
+    private _preprocess;
+    private _buildResult;
+    private _buildBoxes;
+}
+//# sourceMappingURL=detector.d.ts.map

package/dist/tasks/detector.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"detector.d.ts","sourceRoot":"","sources":["../../src/tasks/detector.ts"],"names":[],"mappings":"AAAA;;GAEG;AAIH,OAAO,EAAE,KAAK,WAAW,EAAE,KAAK,iBAAiB,EAAc,MAAM,oBAAoB,CAAC;AAC1F,OAAO,EAAE,KAAK,UAAU,EAAa,MAAM,gBAAgB,CAAC;AAC5D,OAAO,EAAE,KAAK,SAAS,EAAiB,MAAM,cAAc,CAAC;AAQ7D,OAAO,EAAS,gBAAgB,EAAE,MAAM,eAAe,CAAC;AACxD,OAAO,EAAE,UAAU,EAAE,MAAM,WAAW,CAAC;AAOvC;;;;;;;;;GASG;AACH,MAAM,MAAM,YAAY,GAAG,MAAM,CAAC;AAElC,MAAM,WAAW,eAAgB,SAAQ,iBAAiB;IACxD;;;OAGG;IACH,QAAQ,CAAC,IAAI,CAAC,EAAE,YAAY,CAAC;IAC7B,0FAA0F;IAC1F,QAAQ,CAAC,MAAM,CAAC,EAAE,SAAS,CAAC;IAC5B,gEAAgE;IAChE,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,CAAC;IAC7B,gFAAgF;IAChF,QAAQ,CAAC,SAAS,CAAC,EAAE,SAAS,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC/C,uDAAuD;IACvD,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAC;IAChC,yDAAyD;IACzD,QAAQ,CAAC,YAAY,CAAC,EAAE,MAAM,CAAC;IAC/B,8CAA8C;IAC9C,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAC;CACjC;AAED,MAAM,WAAW,sBAAsB;IACrC,iDAAiD;IACjD,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAC;IAChC,0CAA0C;IAC1C,QAAQ,CAAC,YAAY,CAAC,EAAE,MAAM,CAAC;IAC/B;;;OAGG;IACH,QAAQ,CAAC,OAAO,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;CACtC;AAED;;;;;;;;;;;;;;;;;;GAkBG;AACH,qBAAa,QAAS,SAAQ,UAAU;IAGpC,OAAO,CAAC,QAAQ,CAAC,KAAK;IACtB,OAAO,CAAC,QAAQ,CAAC,OAAO;IACxB,OAAO,CAAC,QAAQ,CAAC,MAAM;IACvB,OAAO,CAAC,QAAQ,CAAC,UAAU;IAC3B,OAAO,CAAC,QAAQ,CAAC,cAAc;IAC/B,OAAO,CAAC,QAAQ,CAAC,aAAa;IAC9B,OAAO,CAAC,QAAQ,CAAC,cAAc;IARjC,OAAO;IAaP,yCAAyC;WAC5B,MAAM,CACjB,KAAK,EAAE,WAAW,EAClB,OAAO,GAAE,eAAoB,GAC5B,OAAO,CAAC,QAAQ,CAAC;IAyBpB,+DAA+D;IAC/D,IAAI,IAAI,IAAI,YAAY,CAEvB;IAED,wCAAwC;IACxC,IAAI,MAAM,IAAI,SAAS,MAAM,EAAE,CAE9B;IAED,uEAAuE;IACvE,IAAI,KAAK,IAAI,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,CAE5C;IAED,4CAA4C;IAC5C,IAAI,UAAU,IAAI,MAAM,CAEvB;IAED;;;;;;OAMG;IACG,IAAI,CACR,KAAK,EAAE,UAAU,EACjB,OAAO,GAAE,sBAA2B,GACnC,OAAO,CAAC,gBAAgB,EAAE,CAAC;IAI9B,uCAAuC;IACjC,OAAO,CACX,KAAK,EAAE,UAAU,EACjB,OAAO,GAAE,sBAA2B,GACnC,OAAO,CAAC,gBAAgB,EAAE,CAAC;IAmD9B,OAAO,CAAC,WAAW;IAkBnB,OAAO,CAAC,YAAY;IA4CpB,OAAO,CAAC,WAAW;CAmBpB"}

package/dist/tasks/detector.js

@@ -0,0 +1,189 @@
+/**
+ * Object detection task using anchor-free YOLO ONNX models (v8/v9/v10/v11/v12).
+ */
+import { OrtSession } from "../core/session.js";
+import { loadImage } from "../io/image.js";
+import { resolveLabels } from "../labels.js";
+import { decodeYolo } from "../postprocess/detection.js";
+import { letterbox, toCHW, toFloat32, toFloat32Tensor, } from "../preprocess/image.js";
+import { Boxes, DetectionResults } from "../results.js";
+import { VisionTask } from "./base.js";
+import { RGBImage, } from "../types.js";
+/**
+ * Object detector for anchor-free YOLO ONNX models (v8/v9/v10/v11/v12).
+ *
+ * `predict()` returns `Promise<DetectionResults[]>` (length 1 for a single
+ * image), mirroring Ultralytics' `YOLO("img.jpg")`. Iterate the envelope for
+ * per-instance dataclasses, or use the bulk `boxes` view (`.xyxy`, `.xywh`,
+ * `.xyxyn`, `.xywhn`, `.cls`, `.conf`).
+ *
+ * @example
+ * ```typescript
+ * const det = await Detector.create("/models/yolov8n.onnx");
+ * const results = await det.predict("/images/street.jpg");
+ * const r = results[0];
+ * console.log(r.boxes.xyxy, r.boxes.cls, r.boxes.conf, r.names);
+ * for (const d of r) {
+ *   console.log(d.cls, d.conf, d.box.xyxy);
+ * }
+ * ```
+ */
+export class Detector extends VisionTask {
+    _head;
+    _labels;
+    _names;
+    _inputSize;
+    _confThreshold;
+    _iouThreshold;
+    _maxDetections;
+    constructor(session, _head, _labels, _names, _inputSize, _confThreshold, _iouThreshold, _maxDetections) {
+        super(session);
+        this._head = _head;
+        this._labels = _labels;
+        this._names = _names;
+        this._inputSize = _inputSize;
+        this._confThreshold = _confThreshold;
+        this._iouThreshold = _iouThreshold;
+        this._maxDetections = _maxDetections;
+    }
+    /** Load the model and resolve labels. */
+    static async create(model, options = {}) {
+        const head = options.head ?? "yolo";
+        if (head !== "yolo") {
+            throw new Error(`Unsupported detector head '${head}'. Supported: 'yolo'.`);
+        }
+        const session = await OrtSession.create(model, options);
+        const labels = resolveLabels(options.labels ?? "coco", {
+            numClasses: options.numClasses,
+        });
+        const names = {};
+        for (let i = 0; i < labels.length; i++) {
+            names[i] = labels[i];
+        }
+        return new Detector(session, head, labels, names, options.inputSize ?? [640, 640], options.confThreshold ?? 0.25, options.iouThreshold ?? 0.45, options.maxDetections ?? 300);
+    }
+    /** The decoder family used to interpret the model's output. */
+    get head() {
+        return this._head;
+    }
+    /** Class labels indexed by class id. */
+    get labels() {
+        return this._labels;
+    }
+    /** Class id → class name dict (matches Ultralytics' `model.names`). */
+    get names() {
+        return this._names;
+    }
+    /** Number of classes the model predicts. */
+    get numClasses() {
+        return this._labels.length;
+    }
+    /**
+     * Alias for {@link predict} — call the detector like a torch `nn.Module`.
+     *
+     * Use as `det.call(img)` since JavaScript class instances are not callable;
+     * for direct invocation, prefer `det.predict(img)`. The full
+     * {@link DetectorPredictOptions} (including `classes`) is supported.
+     */
+    async call(image, options = {}) {
+        return this.predict(image, options);
+    }
+    /** Run detection on a single image. */
+    async predict(image, options = {}) {
+        const path = typeof image === "string" ? image : null;
+        const original = await loadImage(image);
+        const { tensor, scale, padLeft, padTop } = this._preprocess(original);
+        const outputs = await this._session.run({ [this._session.inputName]: tensor });
+        const firstOutputName = this._session.outputNames[0];
+        if (firstOutputName === undefined) {
+            throw new Error("Detector model has no outputs.");
+        }
+        const raw = outputs[firstOutputName];
+        if (raw === undefined) {
+            throw new Error(`Detector model output ${firstOutputName} missing from run() result.`);
+        }
+        const decodedAll = decodeYolo(raw.data, raw.dims, {
+            originalWidth: original.width,
+            originalHeight: original.height,
+            padLeft,
+            padTop,
+            scale,
+            confThreshold: options.confThreshold ?? this._confThreshold,
+            iouThreshold: options.iouThreshold ?? this._iouThreshold,
+            maxDetections: this._maxDetections,
+        });
+        const decoded = options.classes !== undefined
+            ? (() => {
+                const allowed = new Set(options.classes);
+                return decodedAll.filter((d) => allowed.has(d.classId));
+            })()
+            : decodedAll;
+        const detections = decoded.map((d) => this._buildResult(original, d.bbox, d.classId, d.confidence));
+        const orig = [original.height, original.width];
+        return [
+            new DetectionResults(this._buildBoxes(detections, orig), detections, this._names, original, orig, path),
+        ];
+    }
+    _preprocess(image) {
+        const [tw, th] = this._inputSize;
+        const lb = letterbox(image, tw, th);
+        const f32 = toFloat32(lb.image);
+        const chw = toCHW(f32, lb.image.width, lb.image.height, 3);
+        return {
+            tensor: toFloat32Tensor(chw, [1, 3, lb.image.height, lb.image.width]),
+            scale: lb.scale,
+            padLeft: lb.padLeft,
+            padTop: lb.padTop,
+        };
+    }
+    _buildResult(original, bbox, classId, confidence) {
+        const [x1, y1, x2, y2] = bbox.asIntXyxy();
+        const cx1 = Math.max(0, x1);
+        const cy1 = Math.max(0, y1);
+        const cx2 = Math.min(original.width, x2);
+        const cy2 = Math.min(original.height, y2);
+        let cropped;
+        if (cx2 > cx1 && cy2 > cy1) {
+            const cw = cx2 - cx1;
+            const ch = cy2 - cy1;
+            const out = new Uint8Array(cw * ch * 3);
+            for (let row = 0; row < ch; row++) {
+                const srcOffset = ((cy1 + row) * original.width + cx1) * 3;
+                out.set(original.data.subarray(srcOffset, srcOffset + cw * 3), row * cw * 3);
+            }
+            cropped = new RGBImage(out, cw, ch);
+        }
+        else {
+            cropped = new RGBImage(new Uint8Array(0), 0, 0);
+        }
+        const className = this._names[classId] ?? `class_${classId}`;
+        return {
+            classId,
+            className,
+            confidence,
+            bbox,
+            cls: classId,
+            name: className,
+            conf: confidence,
+            box: bbox,
+            croppedImage: cropped,
+        };
+    }
+    _buildBoxes(detections, origShape) {
+        const n = detections.length;
+        const xyxy = new Float32Array(n * 4);
+        const cls = new Int32Array(n);
+        const conf = new Float32Array(n);
+        for (let i = 0; i < n; i++) {
+            const d = detections[i];
+            xyxy[i * 4] = d.bbox.x1;
+            xyxy[i * 4 + 1] = d.bbox.y1;
+            xyxy[i * 4 + 2] = d.bbox.x2;
+            xyxy[i * 4 + 3] = d.bbox.y2;
+            cls[i] = d.classId;
+            conf[i] = d.confidence;
+        }
+        return new Boxes(xyxy, cls, conf, origShape);
+    }
+}
+//# sourceMappingURL=detector.js.map

package/dist/tasks/detector.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"detector.js","sourceRoot":"","sources":["../../src/tasks/detector.ts"],"names":[],"mappings":"AAAA;;GAEG;AAIH,OAAO,EAA4C,UAAU,EAAE,MAAM,oBAAoB,CAAC;AAC1F,OAAO,EAAmB,SAAS,EAAE,MAAM,gBAAgB,CAAC;AAC5D,OAAO,EAAkB,aAAa,EAAE,MAAM,cAAc,CAAC;AAC7D,OAAO,EAAE,UAAU,EAAE,MAAM,6BAA6B,CAAC;AACzD,OAAO,EACL,SAAS,EACT,KAAK,EACL,SAAS,EACT,eAAe,GAChB,MAAM,wBAAwB,CAAC;AAChC,OAAO,EAAE,KAAK,EAAE,gBAAgB,EAAE,MAAM,eAAe,CAAC;AACxD,OAAO,EAAE,UAAU,EAAE,MAAM,WAAW,CAAC;AACvC,OAAO,EAGL,QAAQ,GACT,MAAM,aAAa,CAAC;AA8CrB;;;;;;;;;;;;;;;;;;GAkBG;AACH,MAAM,OAAO,QAAS,SAAQ,UAAU;IAGnB;IACA;IACA;IACA;IACA;IACA;IACA;IARnB,YACE,OAAmB,EACF,KAAmB,EACnB,OAA0B,EAC1B,MAAwC,EACxC,UAAqC,EACrC,cAAsB,EACtB,aAAqB,EACrB,cAAsB;QAEvC,KAAK,CAAC,OAAO,CAAC,CAAC;QARE,UAAK,GAAL,KAAK,CAAc;QACnB,YAAO,GAAP,OAAO,CAAmB;QAC1B,WAAM,GAAN,MAAM,CAAkC;QACxC,eAAU,GAAV,UAAU,CAA2B;QACrC,mBAAc,GAAd,cAAc,CAAQ;QACtB,kBAAa,GAAb,aAAa,CAAQ;QACrB,mBAAc,GAAd,cAAc,CAAQ;IAGzC,CAAC;IAED,yCAAyC;IACzC,MAAM,CAAC,KAAK,CAAC,MAAM,CACjB,KAAkB,EAClB,UAA2B,EAAE;QAE7B,MAAM,IAAI,GAAiB,OAAO,CAAC,IAAI,IAAI,MAAM,CAAC;QAClD,IAAI,IAAI,KAAK,MAAM,EAAE,CAAC;YACpB,MAAM,IAAI,KAAK,CAAC,8BAA8B,IAAI,uBAAuB,CAAC,CAAC;QAC7E,CAAC;QACD,MAAM,OAAO,GAAG,MAAM,UAAU,CAAC,MAAM,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC;QACxD,MAAM,MAAM,GAAG,aAAa,CAAC,OAAO,CAAC,MAAM,IAAI,MAAM,EAAE;YACrD,UAAU,EAAE,OAAO,CAAC,UAAU;SAC/B,CAAC,CAAC;QACH,MAAM,KAAK,GAA2B,EAAE,CAAC;QACzC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACvC,KAAK,CAAC,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC,CAAW,CAAC;QACjC,CAAC;QACD,OAAO,IAAI,QAAQ,CACjB,OAAO,EACP,IAAI,EACJ,MAAM,EACN,KAAK,EACL,OAAO,CAAC,SAAS,IAAI,CAAC,GAAG,EAAE,GAAG,CAAC,EAC/B,OAAO,CAAC,aAAa,IAAI,IAAI,EAC7B,OAAO,CAAC,YAAY,IAAI,IAAI,EAC5B,OAAO,CAAC,aAAa,IAAI,GAAG,CAC7B,CAAC;IACJ,CAAC;IAED,+DAA+D;IAC/D,IAAI,IAAI;QACN,OAAO,IAAI,CAAC,KAAK,CAAC;IACpB,CAAC;IAED,wCAAwC;IACxC,IAAI,MAAM;QACR,OAAO,IAAI,CAAC,OAAO,CAAC;IACtB,CAAC;IAED,uEAAuE;IACvE,IAAI,KAAK;QACP,OAAO,IAAI,CAAC,MAAM,CAAC;IACrB,CAAC;IAED,4CAA4C;IAC5C,IAAI,UAAU;QACZ,OAAO,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC;IAC7B,CAAC;IAED;;;;;;OAMG;IACH,KAAK,CAAC,IAAI,CACR,KAAiB,EACjB,UAAkC,EAAE;QAEpC,OAAO,IAAI,CAAC,OAAO,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC;IACtC,CAAC;IAED,uCAAuC;IACvC,KAAK,CAAC,OAAO,CACX,KAAiB,EACjB,UAAkC,EAAE;QAEpC,MAAM,IAAI,GAAG,OAAO,KAAK,KAAK,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC;QACtD,MAAM,QAAQ,GAAG,MAAM,SAAS,CAAC,KAAK,CAAC,CAAC;QACxC,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,OAAO,EAAE,MAAM,EAAE,GAAG,IAAI,CAAC,WAAW,CAAC,QAAQ,CAAC,CAAC;QACtE,MAAM,OAAO,GAAG,MAAM,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,QAAQ,CAAC,SAAS,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC;QAE/E,MAAM,eAAe,GAAG,IAAI,CAAC,QAAQ,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC;QACrD,IAAI,eAAe,KAAK,SAAS,EAAE,CAAC;YAClC,MAAM,IAAI,KAAK,CAAC,gCAAgC,CAAC,CAAC;QACpD,CAAC;QACD,MAAM,GAAG,GAAG,OAAO,CAAC,eAAe,CAAC,CAAC;QACrC,IAAI,GAAG,KAAK,SAAS,EAAE,CAAC;YACtB,MAAM,IAAI,KAAK,CAAC,yBAAyB,eAAe,6BAA6B,CAAC,CAAC;QACzF,CAAC;QAED,MAAM,UAAU,GAAG,UAAU,CAAC,GAAG,CAAC,IAAoB,EAAE,GAAG,CAAC,IAAI,EAAE;YAChE,aAAa,EAAE,QAAQ,CAAC,KAAK;YAC7B,cAAc,EAAE,QAAQ,CAAC,MAAM;YAC/B,OAAO;YACP,MAAM;YACN,KAAK;YACL,aAAa,EAAE,OAAO,CAAC,aAAa,IAAI,IAAI,CAAC,cAAc;YAC3D,YAAY,EAAE,OAAO,CAAC,YAAY,IAAI,IAAI,CAAC,aAAa;YACxD,aAAa,EAAE,IAAI,CAAC,cAAc;SACnC,CAAC,CAAC;QAEH,MAAM,OAAO,GACX,OAAO,CAAC,OAAO,KAAK,SAAS;YAC3B,CAAC,CAAC,CAAC,GAAG,EAAE;gBACJ,MAAM,OAAO,GAAG,IAAI,GAAG,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;gBACzC,OAAO,UAAU,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC;YAC1D,CAAC,CAAC,EAAE;YACN,CAAC,CAAC,UAAU,CAAC;QAEjB,MAAM,UAAU,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CACnC,IAAI,CAAC,YAAY,CAAC,QAAQ,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,UAAU,CAAC,CAC7D,CAAC;QAEF,MAAM,IAAI,GAA8B,CAAC,QAAQ,CAAC,MAAM,EAAE,QAAQ,CAAC,KAAK,CAAC,CAAC;QAC1E,OAAO;YACL,IAAI,gBAAgB,CAClB,IAAI,CAAC,WAAW,CAAC,UAAU,EAAE,IAAI,CAAC,EAClC,UAAU,EACV,IAAI,CAAC,MAAM,EACX,QAAQ,EACR,IAAI,EACJ,IAAI,CACL;SACF,CAAC;IACJ,CAAC;IAEO,WAAW,CAAC,KAAe;QAMjC,MAAM,CAAC,EAAE,EAAE,EAAE,CAAC,GAAG,IAAI,CAAC,UAAU,CAAC;QACjC,MAAM,EAAE,GAAG,SAAS,CAAC,KAAK,EAAE,EAAE,EAAE,EAAE,CAAC,CAAC;QACpC,MAAM,GAAG,GAAG,SAAS,CAAC,EAAE,CAAC,KAAK,CAAC,CAAC;QAChC,MAAM,GAAG,GAAG,KAAK,CAAC,GAAG,EAAE,EAAE,CAAC,KAAK,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC;QAC3D,OAAO;YACL,MAAM,EAAE,eAAe,CAAC,GAAG,EAAE,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,KAAK,CAAC,MAAM,EAAE,EAAE,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;YACrE,KAAK,EAAE,EAAE,CAAC,KAAK;YACf,OAAO,EAAE,EAAE,CAAC,OAAO;YACnB,MAAM,EAAE,EAAE,CAAC,MAAM;SAClB,CAAC;IACJ,CAAC;IAEO,YAAY,CAClB,QAAkB,EAClB,IAAiB,EACjB,OAAe,EACf,UAAkB;QAElB,MAAM,CAAC,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,CAAC,GAAG,IAAI,CAAC,SAAS,EAAE,CAAC;QAC1C,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;QAC5B,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;QAC5B,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,QAAQ,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC;QACzC,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,QAAQ,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC;QAE1C,IAAI,OAAiB,CAAC;QACtB,IAAI,GAAG,GAAG,GAAG,IAAI,GAAG,GAAG,GAAG,EAAE,CAAC;YAC3B,MAAM,EAAE,GAAG,GAAG,GAAG,GAAG,CAAC;YACrB,MAAM,EAAE,GAAG,GAAG,GAAG,GAAG,CAAC;YACrB,MAAM,GAAG,GAAG,IAAI,UAAU,CAAC,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC,CAAC;YACxC,KAAK,IAAI,GAAG,GAAG,CAAC,EAAE,GAAG,GAAG,EAAE,EAAE,GAAG,EAAE,EAAE,CAAC;gBAClC,MAAM,SAAS,GAAG,CAAC,CAAC,GAAG,GAAG,GAAG,CAAC,GAAG,QAAQ,CAAC,KAAK,GAAG,GAAG,CAAC,GAAG,CAAC,CAAC;gBAC3D,GAAG,CAAC,GAAG,CACL,QAAQ,CAAC,IAAI,CAAC,QAAQ,CAAC,SAAS,EAAE,SAAS,GAAG,EAAE,GAAG,CAAC,CAAC,EACrD,GAAG,GAAG,EAAE,GAAG,CAAC,CACb,CAAC;YACJ,CAAC;YACD,OAAO,GAAG,IAAI,QAAQ,CAAC,GAAG,EAAE,EAAE,EAAE,EAAE,CAAC,CAAC;QACtC,CAAC;aAAM,CAAC;YACN,OAAO,GAAG,IAAI,QAAQ,CAAC,IAAI,UAAU,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC;QAClD,CAAC;QAED,MAAM,SAAS,GAAG,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,SAAS,OAAO,EAAE,CAAC;QAE7D,OAAO;YACL,OAAO;YACP,SAAS;YACT,UAAU;YACV,IAAI;YACJ,GAAG,EAAE,OAAO;YACZ,IAAI,EAAE,SAAS;YACf,IAAI,EAAE,UAAU;YAChB,GAAG,EAAE,IAAI;YACT,YAAY,EAAE,OAAO;SACtB,CAAC;IACJ,CAAC;IAEO,WAAW,CACjB,UAAsC,EACtC,SAAoC;QAEpC,MAAM,CAAC,GAAG,UAAU,CAAC,MAAM,CAAC;QAC5B,MAAM,IAAI,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QACrC,MAAM,GAAG,GAAG,IAAI,UAAU,CAAC,CAAC,CAAC,CAAC;QAC9B,MAAM,IAAI,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACjC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC3B,MAAM,CAAC,GAAG,UAAU,CAAC,CAAC,CAAoB,CAAC;YAC3C,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;YACxB,IAAI,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;YAC5B,IAAI,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;YAC5B,IAAI,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;YAC5B,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,OAAO,CAAC;YACnB,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,UAAU,CAAC;QACzB,CAAC;QACD,OAAO,IAAI,KAAK,CAAC,IAAI,EAAE,GAAG,EAAE,IAAI,EAAE,SAAS,CAAC,CAAC;IAC/C,CAAC;CACF"}

package/dist/tasks/segmenter.d.ts

@@ -0,0 +1,105 @@
+/**
+ * Instance-segmentation task using YOLO seg ONNX models (v8-seg / v11-seg / ...).
+ */
+import { type ModelSource, type OrtSessionOptions } from "../core/session.js";
+import { type ImageInput } from "../io/image.js";
+import { type LabelSpec } from "../labels.js";
+import { SegmentationResults } from "../results.js";
+import { VisionTask } from "./base.js";
+/**
+ * Decoder family for the segmentation head.
+ *
+ * - `"yolo-seg"`: YOLO instance-segmentation head with two outputs —
+ *   `[1, 4 + nc + nm, N]` per-anchor predictions plus `[1, nm, mh, mw]`
+ *   prototype masks. Covers YOLOv8-seg, v11-seg, v26-seg.
+ *
+ * The SDK does **not** auto-detect this — the caller is responsible for
+ * picking a head that matches their export.
+ */
+export type SegmenterHead = "yolo-seg";
+export interface SegmenterOptions extends OrtSessionOptions {
+    /**
+     * Decoder family for the segmentation head. Default `"yolo-seg"` covers
+     * YOLOv8-seg/v11-seg/v26-seg.
+     */
+    readonly head?: SegmenterHead;
+    /** Class label spec — see {@link resolveLabels}. Defaults to the COCO 80-class preset. */
+    readonly labels?: LabelSpec;
+    /** Number of classes — used to validate the supplied labels. */
+    readonly numClasses?: number;
+    /** Model input `[width, height]` for letterboxing. Defaults to `[640, 640]`. */
+    readonly inputSize?: readonly [number, number];
+    /** Default minimum class score to keep a candidate. */
+    readonly confThreshold?: number;
+    /** Default IoU threshold for non-maximum suppression. */
+    readonly iouThreshold?: number;
+    /** Maximum number of instances per image. */
+    readonly maxDetections?: number;
+    /** Probability cutoff applied to soft masks. Defaults to `0.5`. */
+    readonly maskThreshold?: number;
+}
+export interface SegmenterPredictOptions {
+    readonly confThreshold?: number;
+    readonly iouThreshold?: number;
+    /**
+     * If set, keep only instances whose `classId` is in this list.
+     * Mirrors Ultralytics' `model.predict(img, classes=[0, 16])`.
+     */
+    readonly classes?: readonly number[];
+}
+/**
+ * Instance segmenter for YOLO seg ONNX models (v8-seg / v11-seg / ...).
+ *
+ * The model is expected to expose two outputs:
+ *
+ * 1. `output0`: `(1, 4 + numClasses + numMaskCoefs, numAnchors)` — per-anchor
+ *    predictions (boxes, class scores, mask coefficients).
+ * 2. `output1`: `(1, numMaskCoefs, maskH, maskW)` — prototype masks.
+ *
+ * `predict()` returns `Promise<SegmentationResults[]>` (length 1 for a
+ * single image), mirroring Ultralytics' API. The envelope exposes:
+ *
+ * - `boxes`: bulk numpy view (`xyxy`, `xywh`, `xyxyn`, `xywhn`, `cls`, `conf`).
+ * - `masks`: per-instance binary masks cropped to each box.
+ * - per-instance {@link SegmentationResult} via iteration.
+ *
+ * @example
+ * ```typescript
+ * const seg = await Segmenter.create("/models/yolov8n-seg.onnx");
+ * const r = (await seg.predict("/images/street.jpg"))[0];
+ * for (const inst of r) {
+ *   console.log(inst.cls, inst.conf, inst.box.xyxy);
+ * }
+ * ```
+ */
+export declare class Segmenter extends VisionTask {
+    private readonly _head;
+    private readonly _labels;
+    private readonly _names;
+    private readonly _inputSize;
+    private readonly _confThreshold;
+    private readonly _iouThreshold;
+    private readonly _maxDetections;
+    private readonly _maskThreshold;
+    private constructor();
+    /** Load the model and resolve labels. */
+    static create(model: ModelSource, options?: SegmenterOptions): Promise<Segmenter>;
+    /** The decoder family used to interpret the model's output. */
+    get head(): SegmenterHead;
+    /** Class labels indexed by class id. */
+    get labels(): readonly string[];
+    /** Class id → class name dict (matches Ultralytics' `model.names`). */
+    get names(): Readonly<Record<number, string>>;
+    /** Number of classes the model predicts. */
+    get numClasses(): number;
+    /** Alias for {@link predict} (parity with PyTorch `nn.Module.__call__`). */
+    call(image: ImageInput, options?: SegmenterPredictOptions): Promise<SegmentationResults[]>;
+    /** Run instance segmentation on a single image. */
+    predict(image: ImageInput, options?: SegmenterPredictOptions): Promise<SegmentationResults[]>;
+    private _preprocess;
+    private _splitOutputs;
+    private _buildResult;
+    private _buildBoxes;
+    private _buildMasks;
+}
+//# sourceMappingURL=segmenter.d.ts.map

package/dist/tasks/segmenter.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"segmenter.d.ts","sourceRoot":"","sources":["../../src/tasks/segmenter.ts"],"names":[],"mappings":"AAAA;;GAEG;AAIH,OAAO,EACL,KAAK,WAAW,EAChB,KAAK,iBAAiB,EAEvB,MAAM,oBAAoB,CAAC;AAC5B,OAAO,EAAE,KAAK,UAAU,EAAa,MAAM,gBAAgB,CAAC;AAC5D,OAAO,EAAE,KAAK,SAAS,EAAiB,MAAM,cAAc,CAAC;AAQ7D,OAAO,EAAgB,mBAAmB,EAAE,MAAM,eAAe,CAAC;AAClE,OAAO,EAAE,UAAU,EAAE,MAAM,WAAW,CAAC;AAQvC;;;;;;;;;GASG;AACH,MAAM,MAAM,aAAa,GAAG,UAAU,CAAC;AAEvC,MAAM,WAAW,gBAAiB,SAAQ,iBAAiB;IACzD;;;OAGG;IACH,QAAQ,CAAC,IAAI,CAAC,EAAE,aAAa,CAAC;IAC9B,0FAA0F;IAC1F,QAAQ,CAAC,MAAM,CAAC,EAAE,SAAS,CAAC;IAC5B,gEAAgE;IAChE,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,CAAC;IAC7B,gFAAgF;IAChF,QAAQ,CAAC,SAAS,CAAC,EAAE,SAAS,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC/C,uDAAuD;IACvD,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAC;IAChC,yDAAyD;IACzD,QAAQ,CAAC,YAAY,CAAC,EAAE,MAAM,CAAC;IAC/B,6CAA6C;IAC7C,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAC;IAChC,mEAAmE;IACnE,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAC;CACjC;AAED,MAAM,WAAW,uBAAuB;IACtC,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAC;IAChC,QAAQ,CAAC,YAAY,CAAC,EAAE,MAAM,CAAC;IAC/B;;;OAGG;IACH,QAAQ,CAAC,OAAO,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;CACtC;AAED;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AACH,qBAAa,SAAU,SAAQ,UAAU;IAGrC,OAAO,CAAC,QAAQ,CAAC,KAAK;IACtB,OAAO,CAAC,QAAQ,CAAC,OAAO;IACxB,OAAO,CAAC,QAAQ,CAAC,MAAM;IACvB,OAAO,CAAC,QAAQ,CAAC,UAAU;IAC3B,OAAO,CAAC,QAAQ,CAAC,cAAc;IAC/B,OAAO,CAAC,QAAQ,CAAC,aAAa;IAC9B,OAAO,CAAC,QAAQ,CAAC,cAAc;IAC/B,OAAO,CAAC,QAAQ,CAAC,cAAc;IATjC,OAAO;IAcP,yCAAyC;WAC5B,MAAM,CACjB,KAAK,EAAE,WAAW,EAClB,OAAO,GAAE,gBAAqB,GAC7B,OAAO,CAAC,SAAS,CAAC;IA0BrB,+DAA+D;IAC/D,IAAI,IAAI,IAAI,aAAa,CAExB;IAED,wCAAwC;IACxC,IAAI,MAAM,IAAI,SAAS,MAAM,EAAE,CAE9B;IAED,uEAAuE;IACvE,IAAI,KAAK,IAAI,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,CAE5C;IAED,4CAA4C;IAC5C,IAAI,UAAU,IAAI,MAAM,CAEvB;IAED,4EAA4E;IACtE,IAAI,CACR,KAAK,EAAE,UAAU,EACjB,OAAO,GAAE,uBAA4B,GACpC,OAAO,CAAC,mBAAmB,EAAE,CAAC;IAIjC,mDAAmD;IAC7C,OAAO,CACX,KAAK,EAAE,UAAU,EACjB,OAAO,GAAE,uBAA4B,GACpC,OAAO,CAAC,mBAAmB,EAAE,CAAC;IAuDjC,OAAO,CAAC,WAAW;IAkBnB,OAAO,CAAC,aAAa;IAyBrB,OAAO,CAAC,YAAY;IAoEpB,OAAO,CAAC,WAAW;IAoBnB,OAAO,CAAC,WAAW;CAkBpB"}

package/dist/tasks/segmenter.js

@@ -0,0 +1,242 @@
+/**
+ * Instance-segmentation task using YOLO seg ONNX models (v8-seg / v11-seg / ...).
+ */
+import { OrtSession, } from "../core/session.js";
+import { loadImage } from "../io/image.js";
+import { resolveLabels } from "../labels.js";
+import { decodeYoloSeg } from "../postprocess/segmentation.js";
+import { letterbox, toCHW, toFloat32, toFloat32Tensor, } from "../preprocess/image.js";
+import { Boxes, Masks, SegmentationResults } from "../results.js";
+import { VisionTask } from "./base.js";
+import { Mask, RGBImage, } from "../types.js";
+/**
+ * Instance segmenter for YOLO seg ONNX models (v8-seg / v11-seg / ...).
+ *
+ * The model is expected to expose two outputs:
+ *
+ * 1. `output0`: `(1, 4 + numClasses + numMaskCoefs, numAnchors)` — per-anchor
+ *    predictions (boxes, class scores, mask coefficients).
+ * 2. `output1`: `(1, numMaskCoefs, maskH, maskW)` — prototype masks.
+ *
+ * `predict()` returns `Promise<SegmentationResults[]>` (length 1 for a
+ * single image), mirroring Ultralytics' API. The envelope exposes:
+ *
+ * - `boxes`: bulk numpy view (`xyxy`, `xywh`, `xyxyn`, `xywhn`, `cls`, `conf`).
+ * - `masks`: per-instance binary masks cropped to each box.
+ * - per-instance {@link SegmentationResult} via iteration.
+ *
+ * @example
+ * ```typescript
+ * const seg = await Segmenter.create("/models/yolov8n-seg.onnx");
+ * const r = (await seg.predict("/images/street.jpg"))[0];
+ * for (const inst of r) {
+ *   console.log(inst.cls, inst.conf, inst.box.xyxy);
+ * }
+ * ```
+ */
+export class Segmenter extends VisionTask {
+    _head;
+    _labels;
+    _names;
+    _inputSize;
+    _confThreshold;
+    _iouThreshold;
+    _maxDetections;
+    _maskThreshold;
+    constructor(session, _head, _labels, _names, _inputSize, _confThreshold, _iouThreshold, _maxDetections, _maskThreshold) {
+        super(session);
+        this._head = _head;
+        this._labels = _labels;
+        this._names = _names;
+        this._inputSize = _inputSize;
+        this._confThreshold = _confThreshold;
+        this._iouThreshold = _iouThreshold;
+        this._maxDetections = _maxDetections;
+        this._maskThreshold = _maskThreshold;
+    }
+    /** Load the model and resolve labels. */
+    static async create(model, options = {}) {
+        const head = options.head ?? "yolo-seg";
+        if (head !== "yolo-seg") {
+            throw new Error(`Unsupported segmenter head '${head}'. Supported: 'yolo-seg'.`);
+        }
+        const session = await OrtSession.create(model, options);
+        const labels = resolveLabels(options.labels ?? "coco", {
+            numClasses: options.numClasses,
+        });
+        const names = {};
+        for (let i = 0; i < labels.length; i++) {
+            names[i] = labels[i];
+        }
+        return new Segmenter(session, head, labels, names, options.inputSize ?? [640, 640], options.confThreshold ?? 0.25, options.iouThreshold ?? 0.45, options.maxDetections ?? 300, options.maskThreshold ?? 0.5);
+    }
+    /** The decoder family used to interpret the model's output. */
+    get head() {
+        return this._head;
+    }
+    /** Class labels indexed by class id. */
+    get labels() {
+        return this._labels;
+    }
+    /** Class id → class name dict (matches Ultralytics' `model.names`). */
+    get names() {
+        return this._names;
+    }
+    /** Number of classes the model predicts. */
+    get numClasses() {
+        return this._labels.length;
+    }
+    /** Alias for {@link predict} (parity with PyTorch `nn.Module.__call__`). */
+    async call(image, options = {}) {
+        return this.predict(image, options);
+    }
+    /** Run instance segmentation on a single image. */
+    async predict(image, options = {}) {
+        const path = typeof image === "string" ? image : null;
+        const original = await loadImage(image);
+        const { tensor, scale, padLeft, padTop } = this._preprocess(original);
+        const outputs = await this._session.run({ [this._session.inputName]: tensor });
+        const { perAnchor, prototypes } = this._splitOutputs(outputs);
+        const decodedAll = decodeYoloSeg(perAnchor.data, perAnchor.dims, prototypes.data, prototypes.dims, {
+            numClasses: this._labels.length,
+            inputWidth: this._inputSize[0],
+            inputHeight: this._inputSize[1],
+            originalWidth: original.width,
+            originalHeight: original.height,
+            padLeft,
+            padTop,
+            scale,
+            confThreshold: options.confThreshold ?? this._confThreshold,
+            iouThreshold: options.iouThreshold ?? this._iouThreshold,
+            maxDetections: this._maxDetections,
+            maskThreshold: this._maskThreshold,
+        });
+        const decoded = options.classes !== undefined
+            ? (() => {
+                const allowed = new Set(options.classes);
+                return decodedAll.filter((d) => allowed.has(d.classId));
+            })()
+            : decodedAll;
+        const detections = decoded.map((d) => this._buildResult(original, d.bbox, d.classId, d.confidence, d.mask));
+        const orig = [original.height, original.width];
+        return [
+            new SegmentationResults(this._buildBoxes(detections, orig), this._buildMasks(detections, orig), detections, this._names, original, orig, path),
+        ];
+    }
+    _preprocess(image) {
+        const [tw, th] = this._inputSize;
+        const lb = letterbox(image, tw, th);
+        const f32 = toFloat32(lb.image);
+        const chw = toCHW(f32, lb.image.width, lb.image.height, 3);
+        return {
+            tensor: toFloat32Tensor(chw, [1, 3, lb.image.height, lb.image.width]),
+            scale: lb.scale,
+            padLeft: lb.padLeft,
+            padTop: lb.padTop,
+        };
+    }
+    _splitOutputs(outputs) {
+        let perAnchor;
+        let prototypes;
+        for (const name of this._session.outputNames) {
+            const t = outputs[name];
+            if (t === undefined)
+                continue;
+            if (t.dims.length === 3 && perAnchor === undefined) {
+                perAnchor = t;
+            }
+            else if (t.dims.length === 4 && prototypes === undefined) {
+                prototypes = t;
+            }
+        }
+        if (perAnchor === undefined || prototypes === undefined) {
+            const shapes = this._session.outputNames.map((n) => `${n}: ${JSON.stringify(outputs[n]?.dims ?? [])}`);
+            throw new Error(`Segmenter expected one 3-D and one 4-D output, got [${shapes.join(", ")}].`);
+        }
+        return { perAnchor, prototypes };
+    }
+    _buildResult(original, bbox, classId, confidence, mask) {
+        const [x1, y1, x2, y2] = bbox.asIntXyxy();
+        const cx1 = Math.max(0, x1);
+        const cy1 = Math.max(0, y1);
+        const cx2 = Math.min(original.width, x2);
+        const cy2 = Math.min(original.height, y2);
+        let segmentedImage;
+        let finalMask = mask;
+        if (cx2 > cx1 && cy2 > cy1 && mask.data.length > 0) {
+            const cropW = cx2 - cx1;
+            const cropH = cy2 - cy1;
+            const mw = Math.min(mask.width, cropW);
+            const mh = Math.min(mask.height, cropH);
+            const segData = new Uint8Array(mw * mh * 3);
+            for (let row = 0; row < mh; row++) {
+                const srcRowOffset = ((cy1 + row) * original.width + cx1) * 3;
+                const dstRowOffset = row * mw * 3;
+                const maskRowOffset = row * mask.width;
+                for (let col = 0; col < mw; col++) {
+                    const m = mask.data[maskRowOffset + col];
+                    if (m !== 0) {
+                        const s = srcRowOffset + col * 3;
+                        const d = dstRowOffset + col * 3;
+                        segData[d] = original.data[s];
+                        segData[d + 1] = original.data[s + 1];
+                        segData[d + 2] = original.data[s + 2];
+                    }
+                }
+            }
+            segmentedImage = new RGBImage(segData, mw, mh);
+            if (mw !== mask.width || mh !== mask.height) {
+                const trimmed = new Uint8Array(mw * mh);
+                for (let row = 0; row < mh; row++) {
+                    trimmed.set(mask.data.subarray(row * mask.width, row * mask.width + mw), row * mw);
+                }
+                finalMask = new Mask(trimmed, mw, mh);
+            }
+        }
+        else {
+            finalMask = new Mask(new Uint8Array(0), 0, 0);
+            segmentedImage = new RGBImage(new Uint8Array(0), 0, 0);
+        }
+        const className = this._names[classId] ?? `class_${classId}`;
+        return {
+            classId,
+            className,
+            confidence,
+            bbox,
+            cls: classId,
+            name: className,
+            conf: confidence,
+            box: bbox,
+            mask: finalMask,
+            segmentedImage,
+        };
+    }
+    _buildBoxes(detections, origShape) {
+        const n = detections.length;
+        const xyxy = new Float32Array(n * 4);
+        const cls = new Int32Array(n);
+        const conf = new Float32Array(n);
+        for (let i = 0; i < n; i++) {
+            const d = detections[i];
+            xyxy[i * 4] = d.bbox.x1;
+            xyxy[i * 4 + 1] = d.bbox.y1;
+            xyxy[i * 4 + 2] = d.bbox.x2;
+            xyxy[i * 4 + 3] = d.bbox.y2;
+            cls[i] = d.classId;
+            conf[i] = d.confidence;
+        }
+        return new Boxes(xyxy, cls, conf, origShape);
+    }
+    _buildMasks(detections, origShape) {
+        const xyxy = new Float32Array(detections.length * 4);
+        for (let i = 0; i < detections.length; i++) {
+            const d = detections[i];
+            xyxy[i * 4] = d.bbox.x1;
+            xyxy[i * 4 + 1] = d.bbox.y1;
+            xyxy[i * 4 + 2] = d.bbox.x2;
+            xyxy[i * 4 + 3] = d.bbox.y2;
+        }
+        return new Masks(detections.map((d) => d.mask), xyxy, origShape);
+    }
+}
+//# sourceMappingURL=segmenter.js.map