@camstack/addon-vision 0.1.2 → 0.1.3

package/dist/chunk-ZK7P3TZN.mjs

@@ -0,0 +1,286 @@
+import {
+  require_yolo
+} from "./chunk-DE7I3VHO.mjs";
+import {
+  require_object_detection_models
+} from "./chunk-MGT6RUVX.mjs";
+import {
+  __commonJS,
+  __require
+} from "./chunk-3IIFBJCD.mjs";
+
+// src/catalogs/segmentation-models.js
+var require_segmentation_models = __commonJS({
+  "src/catalogs/segmentation-models.js"(exports) {
+    "use strict";
+    Object.defineProperty(exports, "__esModule", { value: true });
+    exports.SEGMENTATION_MODELS = void 0;
+    var types_1 = __require("@camstack/types");
+    var object_detection_models_js_1 = require_object_detection_models();
+    var HF_REPO = "camstack/camstack-models";
+    exports.SEGMENTATION_MODELS = [
+      // ── YOLO11-seg ───────────────────────────────────────────────
+      {
+        id: "yolo11n-seg",
+        name: "YOLO11 Nano Segmentation",
+        description: "YOLO11 Nano \u2014 fastest, smallest YOLO11 instance segmentation model",
+        inputSize: { width: 640, height: 640 },
+        labels: types_1.COCO_80_LABELS,
+        formats: {
+          onnx: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolo11-seg/onnx/camstack-yolo11n-seg.onnx"),
+            sizeMB: 11
+          },
+          coreml: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolo11-seg/coreml/camstack-yolo11n-seg.mlpackage"),
+            sizeMB: 6,
+            isDirectory: true,
+            files: object_detection_models_js_1.MLPACKAGE_FILES,
+            runtimes: ["python"]
+          },
+          openvino: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolo11-seg/openvino/camstack-yolo11n-seg.xml"),
+            sizeMB: 6,
+            runtimes: ["python"]
+          }
+        }
+      },
+      {
+        id: "yolo11s-seg",
+        name: "YOLO11 Small Segmentation",
+        description: "YOLO11 Small \u2014 balanced speed and accuracy for instance segmentation",
+        inputSize: { width: 640, height: 640 },
+        labels: types_1.COCO_80_LABELS,
+        formats: {
+          onnx: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolo11-seg/onnx/camstack-yolo11s-seg.onnx"),
+            sizeMB: 39
+          },
+          coreml: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolo11-seg/coreml/camstack-yolo11s-seg.mlpackage"),
+            sizeMB: 20,
+            isDirectory: true,
+            files: object_detection_models_js_1.MLPACKAGE_FILES,
+            runtimes: ["python"]
+          },
+          openvino: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolo11-seg/openvino/camstack-yolo11s-seg.xml"),
+            sizeMB: 20,
+            runtimes: ["python"]
+          }
+        }
+      },
+      {
+        id: "yolo11m-seg",
+        name: "YOLO11 Medium Segmentation",
+        description: "YOLO11 Medium \u2014 higher accuracy instance segmentation",
+        inputSize: { width: 640, height: 640 },
+        labels: types_1.COCO_80_LABELS,
+        formats: {
+          onnx: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolo11-seg/onnx/camstack-yolo11m-seg.onnx"),
+            sizeMB: 86
+          },
+          coreml: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolo11-seg/coreml/camstack-yolo11m-seg.mlpackage"),
+            sizeMB: 43,
+            isDirectory: true,
+            files: object_detection_models_js_1.MLPACKAGE_FILES,
+            runtimes: ["python"]
+          },
+          openvino: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolo11-seg/openvino/camstack-yolo11m-seg.xml"),
+            sizeMB: 43,
+            runtimes: ["python"]
+          }
+        }
+      },
+      // ── YOLOv8-seg — CoreML supported ─────────────────────────────
+      {
+        id: "yolov8n-seg",
+        name: "YOLOv8 Nano Segmentation",
+        description: "YOLOv8 Nano \u2014 fastest, smallest YOLOv8 instance segmentation model",
+        inputSize: { width: 640, height: 640 },
+        labels: types_1.COCO_80_LABELS,
+        formats: {
+          onnx: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolov8-seg/onnx/camstack-yolov8n-seg.onnx"),
+            sizeMB: 13
+          },
+          coreml: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolov8-seg/coreml/camstack-yolov8n-seg.mlpackage"),
+            sizeMB: 7,
+            isDirectory: true,
+            files: object_detection_models_js_1.MLPACKAGE_FILES,
+            runtimes: ["python"]
+          },
+          openvino: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolov8-seg/openvino/camstack-yolov8n-seg.xml"),
+            sizeMB: 7,
+            runtimes: ["python"]
+          }
+        }
+      },
+      {
+        id: "yolov8s-seg",
+        name: "YOLOv8 Small Segmentation",
+        description: "YOLOv8 Small \u2014 balanced speed and accuracy for instance segmentation",
+        inputSize: { width: 640, height: 640 },
+        labels: types_1.COCO_80_LABELS,
+        formats: {
+          onnx: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolov8-seg/onnx/camstack-yolov8s-seg.onnx"),
+            sizeMB: 45
+          },
+          coreml: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolov8-seg/coreml/camstack-yolov8s-seg.mlpackage"),
+            sizeMB: 23,
+            isDirectory: true,
+            files: object_detection_models_js_1.MLPACKAGE_FILES,
+            runtimes: ["python"]
+          },
+          openvino: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolov8-seg/openvino/camstack-yolov8s-seg.xml"),
+            sizeMB: 23,
+            runtimes: ["python"]
+          }
+        }
+      },
+      {
+        id: "yolov8m-seg",
+        name: "YOLOv8 Medium Segmentation",
+        description: "YOLOv8 Medium \u2014 higher accuracy instance segmentation",
+        inputSize: { width: 640, height: 640 },
+        labels: types_1.COCO_80_LABELS,
+        formats: {
+          onnx: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolov8-seg/onnx/camstack-yolov8m-seg.onnx"),
+            sizeMB: 104
+          },
+          coreml: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolov8-seg/coreml/camstack-yolov8m-seg.mlpackage"),
+            sizeMB: 52,
+            isDirectory: true,
+            files: object_detection_models_js_1.MLPACKAGE_FILES,
+            runtimes: ["python"]
+          },
+          openvino: {
+            url: (0, types_1.hfModelUrl)(HF_REPO, "segmentation/yolov8-seg/openvino/camstack-yolov8m-seg.xml"),
+            sizeMB: 53,
+            runtimes: ["python"]
+          }
+        }
+      }
+    ];
+  }
+});
+
+// src/shared/postprocess/yolo-seg.js
+var require_yolo_seg = __commonJS({
+  "src/shared/postprocess/yolo-seg.js"(exports) {
+    "use strict";
+    Object.defineProperty(exports, "__esModule", { value: true });
+    exports.sigmoid = sigmoid;
+    exports.computeRawMask = computeRawMask;
+    exports.cropAndThresholdMask = cropAndThresholdMask;
+    exports.yoloSegPostprocess = yoloSegPostprocess;
+    var yolo_js_1 = require_yolo();
+    function sigmoid(x) {
+      return 1 / (1 + Math.exp(-x));
+    }
+    function computeRawMask(coeffs, protos, numMaskCoeffs, maskH, maskW) {
+      const maskSize = maskH * maskW;
+      const rawMask = new Float32Array(maskSize);
+      for (let px = 0; px < maskSize; px++) {
+        let val = 0;
+        for (let k = 0; k < numMaskCoeffs; k++) {
+          val += (coeffs[k] ?? 0) * (protos[k * maskSize + px] ?? 0);
+        }
+        rawMask[px] = sigmoid(val);
+      }
+      return rawMask;
+    }
+    function cropAndThresholdMask(rawMask, maskH, maskW, bbox, maskThreshold, maskScale) {
+      const cropX1 = Math.max(0, Math.floor(bbox.x * maskScale));
+      const cropY1 = Math.max(0, Math.floor(bbox.y * maskScale));
+      const cropX2 = Math.min(maskW, Math.ceil((bbox.x + bbox.w) * maskScale));
+      const cropY2 = Math.min(maskH, Math.ceil((bbox.y + bbox.h) * maskScale));
+      const cropW = Math.max(1, cropX2 - cropX1);
+      const cropH = Math.max(1, cropY2 - cropY1);
+      const data = new Uint8Array(cropW * cropH);
+      for (let row = 0; row < cropH; row++) {
+        const srcRow = cropY1 + row;
+        for (let col = 0; col < cropW; col++) {
+          const srcCol = cropX1 + col;
+          const srcIdx = srcRow * maskW + srcCol;
+          data[row * cropW + col] = (rawMask[srcIdx] ?? 0) > maskThreshold ? 255 : 0;
+        }
+      }
+      return { data, width: cropW, height: cropH };
+    }
+    function yoloSegPostprocess(segOutput, options) {
+      const { detectionOutput, protoOutput, numClasses, numBoxes, numMaskCoeffs, maskHeight, maskWidth } = segOutput;
+      const { confidence, iouThreshold, labels, scale, padX, padY, originalWidth, originalHeight, maskThreshold = 0.5 } = options;
+      const yoloInputSize = 640;
+      const maskScale = maskHeight / yoloInputSize;
+      const candidates = [];
+      for (let i = 0; i < numBoxes; i++) {
+        const cx = detectionOutput[0 * numBoxes + i] ?? 0;
+        const cy = detectionOutput[1 * numBoxes + i] ?? 0;
+        const w = detectionOutput[2 * numBoxes + i] ?? 0;
+        const h = detectionOutput[3 * numBoxes + i] ?? 0;
+        let bestScore = -Infinity;
+        let bestClass = 0;
+        for (let j = 0; j < numClasses; j++) {
+          const score = detectionOutput[(4 + j) * numBoxes + i] ?? 0;
+          if (score > bestScore) {
+            bestScore = score;
+            bestClass = j;
+          }
+        }
+        if (bestScore < confidence)
+          continue;
+        const bbox = {
+          x: cx - w / 2,
+          y: cy - h / 2,
+          w,
+          h
+        };
+        const coeffs = new Float32Array(numMaskCoeffs);
+        for (let k = 0; k < numMaskCoeffs; k++) {
+          coeffs[k] = detectionOutput[(4 + numClasses + k) * numBoxes + i] ?? 0;
+        }
+        candidates.push({ bbox, score: bestScore, classIdx: bestClass, coeffs });
+      }
+      if (candidates.length === 0)
+        return [];
+      const keptIndices = (0, yolo_js_1.nms)(candidates, iouThreshold);
+      return keptIndices.map((idx) => {
+        const { bbox, score, classIdx, coeffs } = candidates[idx];
+        const label = labels[classIdx] ?? String(classIdx);
+        const x = Math.max(0, Math.min(originalWidth, (bbox.x - padX) / scale));
+        const y = Math.max(0, Math.min(originalHeight, (bbox.y - padY) / scale));
+        const x2 = Math.max(0, Math.min(originalWidth, (bbox.x + bbox.w - padX) / scale));
+        const y2 = Math.max(0, Math.min(originalHeight, (bbox.y + bbox.h - padY) / scale));
+        const finalBbox = { x, y, w: x2 - x, h: y2 - y };
+        const rawMask = computeRawMask(coeffs, protoOutput, numMaskCoeffs, maskHeight, maskWidth);
+        const { data: maskData, width: mW, height: mH } = cropAndThresholdMask(rawMask, maskHeight, maskWidth, bbox, maskThreshold, maskScale);
+        return {
+          class: label,
+          originalClass: label,
+          score,
+          bbox: finalBbox,
+          mask: maskData,
+          maskWidth: mW,
+          maskHeight: mH
+        };
+      });
+    }
+  }
+});
+
+export {
+  require_segmentation_models,
+  require_yolo_seg
+};
+//# sourceMappingURL=chunk-ZK7P3TZN.mjs.map

package/dist/chunk-ZK7P3TZN.mjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/catalogs/segmentation-models.ts","../src/shared/postprocess/yolo-seg.ts"],"sourcesContent":["import type { ModelCatalogEntry } from '@camstack/types'\nimport { hfModelUrl, COCO_80_LABELS } from '@camstack/types'\nimport { MLPACKAGE_FILES } from './object-detection-models.js'\n\nconst HF_REPO = 'camstack/camstack-models'\n\nexport const SEGMENTATION_MODELS: readonly ModelCatalogEntry[] = [\n  // ── YOLO11-seg ───────────────────────────────────────────────\n  {\n    id: 'yolo11n-seg',\n    name: 'YOLO11 Nano Segmentation',\n    description: 'YOLO11 Nano — fastest, smallest YOLO11 instance segmentation model',\n    inputSize: { width: 640, height: 640 },\n    labels: COCO_80_LABELS,\n    formats: {\n      onnx: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolo11-seg/onnx/camstack-yolo11n-seg.onnx'),\n        sizeMB: 11,\n      },\n      coreml: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolo11-seg/coreml/camstack-yolo11n-seg.mlpackage'),\n        sizeMB: 6,\n        isDirectory: true,\n        files: MLPACKAGE_FILES,\n        runtimes: ['python'],\n      },\n      openvino: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolo11-seg/openvino/camstack-yolo11n-seg.xml'),\n        sizeMB: 6,\n        runtimes: ['python'],\n      },\n    },\n  },\n  {\n    id: 'yolo11s-seg',\n    name: 'YOLO11 Small Segmentation',\n    description: 'YOLO11 Small — balanced speed and accuracy for instance segmentation',\n    inputSize: { width: 640, height: 640 },\n    labels: COCO_80_LABELS,\n    formats: {\n      onnx: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolo11-seg/onnx/camstack-yolo11s-seg.onnx'),\n        sizeMB: 39,\n      },\n      coreml: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolo11-seg/coreml/camstack-yolo11s-seg.mlpackage'),\n        sizeMB: 20,\n        isDirectory: true,\n        files: MLPACKAGE_FILES,\n        runtimes: ['python'],\n      },\n      openvino: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolo11-seg/openvino/camstack-yolo11s-seg.xml'),\n        sizeMB: 20,\n        runtimes: ['python'],\n      },\n    },\n  },\n  {\n    id: 'yolo11m-seg',\n    name: 'YOLO11 Medium Segmentation',\n    description: 'YOLO11 Medium — higher accuracy instance segmentation',\n    inputSize: { width: 640, height: 640 },\n    labels: COCO_80_LABELS,\n    formats: {\n      onnx: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolo11-seg/onnx/camstack-yolo11m-seg.onnx'),\n        sizeMB: 86,\n      },\n      coreml: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolo11-seg/coreml/camstack-yolo11m-seg.mlpackage'),\n        sizeMB: 43,\n        isDirectory: true,\n        files: MLPACKAGE_FILES,\n        runtimes: ['python'],\n      },\n      openvino: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolo11-seg/openvino/camstack-yolo11m-seg.xml'),\n        sizeMB: 43,\n        runtimes: ['python'],\n      },\n    },\n  },\n\n  // ── YOLOv8-seg — CoreML supported ─────────────────────────────\n  {\n    id: 'yolov8n-seg',\n    name: 'YOLOv8 Nano Segmentation',\n    description: 'YOLOv8 Nano — fastest, smallest YOLOv8 instance segmentation model',\n    inputSize: { width: 640, height: 640 },\n    labels: COCO_80_LABELS,\n    formats: {\n      onnx: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolov8-seg/onnx/camstack-yolov8n-seg.onnx'),\n        sizeMB: 13,\n      },\n      coreml: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolov8-seg/coreml/camstack-yolov8n-seg.mlpackage'),\n        sizeMB: 7,\n        isDirectory: true,\n        files: MLPACKAGE_FILES,\n        runtimes: ['python'],\n      },\n      openvino: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolov8-seg/openvino/camstack-yolov8n-seg.xml'),\n        sizeMB: 7,\n        runtimes: ['python'],\n      },\n    },\n  },\n  {\n    id: 'yolov8s-seg',\n    name: 'YOLOv8 Small Segmentation',\n    description: 'YOLOv8 Small — balanced speed and accuracy for instance segmentation',\n    inputSize: { width: 640, height: 640 },\n    labels: COCO_80_LABELS,\n    formats: {\n      onnx: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolov8-seg/onnx/camstack-yolov8s-seg.onnx'),\n        sizeMB: 45,\n      },\n      coreml: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolov8-seg/coreml/camstack-yolov8s-seg.mlpackage'),\n        sizeMB: 23,\n        isDirectory: true,\n        files: MLPACKAGE_FILES,\n        runtimes: ['python'],\n      },\n      openvino: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolov8-seg/openvino/camstack-yolov8s-seg.xml'),\n        sizeMB: 23,\n        runtimes: ['python'],\n      },\n    },\n  },\n  {\n    id: 'yolov8m-seg',\n    name: 'YOLOv8 Medium Segmentation',\n    description: 'YOLOv8 Medium — higher accuracy instance segmentation',\n    inputSize: { width: 640, height: 640 },\n    labels: COCO_80_LABELS,\n    formats: {\n      onnx: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolov8-seg/onnx/camstack-yolov8m-seg.onnx'),\n        sizeMB: 104,\n      },\n      coreml: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolov8-seg/coreml/camstack-yolov8m-seg.mlpackage'),\n        sizeMB: 52,\n        isDirectory: true,\n        files: MLPACKAGE_FILES,\n        runtimes: ['python'],\n      },\n      openvino: {\n        url: hfModelUrl(HF_REPO, 'segmentation/yolov8-seg/openvino/camstack-yolov8m-seg.xml'),\n        sizeMB: 53,\n        runtimes: ['python'],\n      },\n    },\n  },\n] as const\n","import type { SpatialDetection, BoundingBox } from '@camstack/types'\nimport { nms, iou as _iou } from './yolo.js'\nimport type { YoloPostprocessOptions } from './yolo.js'\n\nexport interface YoloSegOutput {\n  /** Detection output: [1, 4 + numClasses + numMaskCoeffs, numBoxes] (row-major, batch dim stripped) */\n  readonly detectionOutput: Float32Array\n  /** Prototype masks: [1, numMaskCoeffs, maskH, maskW] (batch dim stripped) */\n  readonly protoOutput: Float32Array\n  readonly numClasses: number\n  readonly numBoxes: number\n  /** Number of mask prototype coefficients per detection (typically 32) */\n  readonly numMaskCoeffs: number\n  /** Height of prototype mask grid (typically 160) */\n  readonly maskHeight: number\n  /** Width of prototype mask grid (typically 160) */\n  readonly maskWidth: number\n}\n\nexport interface YoloSegPostprocessOptions extends YoloPostprocessOptions {\n  /** Binary mask threshold — pixels above this value become 255, others 0 (default: 0.5) */\n  readonly maskThreshold?: number\n}\n\n/** Sigmoid activation: σ(x) = 1 / (1 + e^-x) */\nexport function sigmoid(x: number): number {\n  return 1 / (1 + Math.exp(-x))\n}\n\n/**\n * Compute instance segmentation mask for a single detection.\n *\n * @param coeffs   [numMaskCoeffs] mask coefficients for this detection\n * @param protos   [numMaskCoeffs × maskH × maskW] prototype masks (flat, row-major)\n * @param numMaskCoeffs  number of prototype channels (e.g. 32)\n * @param maskH    prototype grid height (e.g. 160)\n * @param maskW    prototype grid width  (e.g. 160)\n * @returns        [maskH × maskW] sigmoid-activated raw mask values in [0, 1]\n */\nexport function computeRawMask(\n  coeffs: Float32Array,\n  protos: Float32Array,\n  numMaskCoeffs: number,\n  maskH: number,\n  maskW: number,\n): Float32Array {\n  const maskSize = maskH * maskW\n  const rawMask = new Float32Array(maskSize)\n\n  // mask_raw[y * maskW + x] = sum_k( coeffs[k] * protos[k * maskSize + y * maskW + x] )\n  // Then apply sigmoid element-wise.\n  for (let px = 0; px < maskSize; px++) {\n    let val = 0\n    for (let k = 0; k < numMaskCoeffs; k++) {\n      val += (coeffs[k] ?? 0) * (protos[k * maskSize + px] ?? 0)\n    }\n    rawMask[px] = sigmoid(val)\n  }\n\n  return rawMask\n}\n\n/**\n * Crop a flat [maskH × maskW] mask to a bounding box region (in mask coordinate space),\n * threshold to binary, and return the cropped Uint8Array along with its dimensions.\n *\n * @param rawMask       [maskH × maskW] sigmoid-activated values in [0, 1]\n * @param maskH         full mask grid height\n * @param maskW         full mask grid width\n * @param bbox          detection bounding box in the inference input space (e.g. 640×640 letterboxed)\n * @param maskThreshold threshold above which a pixel is considered foreground (default 0.5)\n * @param maskScale     ratio of mask resolution to inference resolution (e.g. 160/640 = 0.25)\n * @returns             { data: Uint8Array, width, height } of the cropped binary mask\n */\nexport function cropAndThresholdMask(\n  rawMask: Float32Array,\n  maskH: number,\n  maskW: number,\n  bbox: BoundingBox,\n  maskThreshold: number,\n  maskScale: number,\n): { readonly data: Uint8Array; readonly width: number; readonly height: number } {\n  const cropX1 = Math.max(0, Math.floor(bbox.x * maskScale))\n  const cropY1 = Math.max(0, Math.floor(bbox.y * maskScale))\n  const cropX2 = Math.min(maskW, Math.ceil((bbox.x + bbox.w) * maskScale))\n  const cropY2 = Math.min(maskH, Math.ceil((bbox.y + bbox.h) * maskScale))\n\n  const cropW = Math.max(1, cropX2 - cropX1)\n  const cropH = Math.max(1, cropY2 - cropY1)\n\n  const data = new Uint8Array(cropW * cropH)\n\n  for (let row = 0; row < cropH; row++) {\n    const srcRow = cropY1 + row\n    for (let col = 0; col < cropW; col++) {\n      const srcCol = cropX1 + col\n      const srcIdx = srcRow * maskW + srcCol\n      data[row * cropW + col] = (rawMask[srcIdx] ?? 0) > maskThreshold ? 255 : 0\n    }\n  }\n\n  return { data, width: cropW, height: cropH }\n}\n\n/**\n * YOLO-seg postprocessing: run detection filtering + NMS, then decode instance masks.\n *\n * The seg detection tensor has layout [1, 4 + numClasses + numMaskCoeffs, numBoxes]:\n *   rows 0–3            : cx, cy, w, h\n *   rows 4 to 4+C-1     : class scores\n *   rows 4+C to end     : mask coefficients (32)\n *\n * Steps:\n *   1. Filter boxes by confidence threshold, collecting mask coefficients per candidate.\n *   2. Run NMS across all candidates.\n *   3. For each surviving detection, compute the instance mask and attach it.\n */\nexport function yoloSegPostprocess(\n  segOutput: YoloSegOutput,\n  options: YoloSegPostprocessOptions,\n): SpatialDetection[] {\n  const {\n    detectionOutput,\n    protoOutput,\n    numClasses,\n    numBoxes,\n    numMaskCoeffs,\n    maskHeight,\n    maskWidth,\n  } = segOutput\n\n  const {\n    confidence,\n    iouThreshold,\n    labels,\n    scale,\n    padX,\n    padY,\n    originalWidth,\n    originalHeight,\n    maskThreshold = 0.5,\n  } = options\n\n  // The mask resolution to inference resolution scale factor (e.g. 160/640 = 0.25)\n  // We derive the inference size from scale/pad context: the letterboxed input is typically 640×640\n  // but we receive coordinates in that space already, so we use maskHeight / inferH.\n  // Since all YOLO-seg models use 640×640 input and 160×160 proto, the ratio is always 160/640.\n  // We compute it generically from the known prototype dims + fixed YOLO input of 640.\n  const yoloInputSize = 640\n  const maskScale = maskHeight / yoloInputSize\n\n  interface Candidate {\n    readonly bbox: BoundingBox\n    readonly score: number\n    readonly classIdx: number\n    readonly coeffs: Float32Array\n  }\n\n  const candidates: Candidate[] = []\n\n  for (let i = 0; i < numBoxes; i++) {\n    const cx = detectionOutput[0 * numBoxes + i] ?? 0\n    const cy = detectionOutput[1 * numBoxes + i] ?? 0\n    const w = detectionOutput[2 * numBoxes + i] ?? 0\n    const h = detectionOutput[3 * numBoxes + i] ?? 0\n\n    let bestScore = -Infinity\n    let bestClass = 0\n\n    for (let j = 0; j < numClasses; j++) {\n      const score = detectionOutput[(4 + j) * numBoxes + i] ?? 0\n      if (score > bestScore) {\n        bestScore = score\n        bestClass = j\n      }\n    }\n\n    if (bestScore < confidence) continue\n\n    const bbox: BoundingBox = {\n      x: cx - w / 2,\n      y: cy - h / 2,\n      w,\n      h,\n    }\n\n    // Extract mask coefficients for this box\n    const coeffs = new Float32Array(numMaskCoeffs)\n    for (let k = 0; k < numMaskCoeffs; k++) {\n      coeffs[k] = detectionOutput[(4 + numClasses + k) * numBoxes + i] ?? 0\n    }\n\n    candidates.push({ bbox, score: bestScore, classIdx: bestClass, coeffs })\n  }\n\n  if (candidates.length === 0) return []\n\n  const keptIndices = nms(candidates, iouThreshold)\n\n  return keptIndices.map((idx) => {\n    const { bbox, score, classIdx, coeffs } = candidates[idx]!\n    const label = labels[classIdx] ?? String(classIdx)\n\n    // Scale bbox back from letterbox space to original image coords\n    const x = Math.max(0, Math.min(originalWidth, (bbox.x - padX) / scale))\n    const y = Math.max(0, Math.min(originalHeight, (bbox.y - padY) / scale))\n    const x2 = Math.max(0, Math.min(originalWidth, (bbox.x + bbox.w - padX) / scale))\n    const y2 = Math.max(0, Math.min(originalHeight, (bbox.y + bbox.h - padY) / scale))\n    const finalBbox: BoundingBox = { x, y, w: x2 - x, h: y2 - y }\n\n    // Decode instance mask using the letterbox-space bbox (before coord scaling)\n    const rawMask = computeRawMask(coeffs, protoOutput, numMaskCoeffs, maskHeight, maskWidth)\n    const { data: maskData, width: mW, height: mH } = cropAndThresholdMask(\n      rawMask,\n      maskHeight,\n      maskWidth,\n      bbox,\n      maskThreshold,\n      maskScale,\n    )\n\n    return {\n      class: label,\n      originalClass: label,\n      score,\n      bbox: finalBbox,\n      mask: maskData,\n      maskWidth: mW,\n      maskHeight: mH,\n    } satisfies SpatialDetection\n  })\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AACA,QAAA,UAAA,UAAA,iBAAA;AACA,QAAA,+BAAA;AAEA,QAAM,UAAU;AAEH,YAAA,sBAAoD;;MAE/D;QACE,IAAI;QACJ,MAAM;QACN,aAAa;QACb,WAAW,EAAE,OAAO,KAAK,QAAQ,IAAG;QACpC,QAAQ,QAAA;QACR,SAAS;UACP,MAAM;YACJ,MAAK,GAAA,QAAA,YAAW,SAAS,wDAAwD;YACjF,QAAQ;;UAEV,QAAQ;YACN,MAAK,GAAA,QAAA,YAAW,SAAS,+DAA+D;YACxF,QAAQ;YACR,aAAa;YACb,OAAO,6BAAA;YACP,UAAU,CAAC,QAAQ;;UAErB,UAAU;YACR,MAAK,GAAA,QAAA,YAAW,SAAS,2DAA2D;YACpF,QAAQ;YACR,UAAU,CAAC,QAAQ;;;;MAIzB;QACE,IAAI;QACJ,MAAM;QACN,aAAa;QACb,WAAW,EAAE,OAAO,KAAK,QAAQ,IAAG;QACpC,QAAQ,QAAA;QACR,SAAS;UACP,MAAM;YACJ,MAAK,GAAA,QAAA,YAAW,SAAS,wDAAwD;YACjF,QAAQ;;UAEV,QAAQ;YACN,MAAK,GAAA,QAAA,YAAW,SAAS,+DAA+D;YACxF,QAAQ;YACR,aAAa;YACb,OAAO,6BAAA;YACP,UAAU,CAAC,QAAQ;;UAErB,UAAU;YACR,MAAK,GAAA,QAAA,YAAW,SAAS,2DAA2D;YACpF,QAAQ;YACR,UAAU,CAAC,QAAQ;;;;MAIzB;QACE,IAAI;QACJ,MAAM;QACN,aAAa;QACb,WAAW,EAAE,OAAO,KAAK,QAAQ,IAAG;QACpC,QAAQ,QAAA;QACR,SAAS;UACP,MAAM;YACJ,MAAK,GAAA,QAAA,YAAW,SAAS,wDAAwD;YACjF,QAAQ;;UAEV,QAAQ;YACN,MAAK,GAAA,QAAA,YAAW,SAAS,+DAA+D;YACxF,QAAQ;YACR,aAAa;YACb,OAAO,6BAAA;YACP,UAAU,CAAC,QAAQ;;UAErB,UAAU;YACR,MAAK,GAAA,QAAA,YAAW,SAAS,2DAA2D;YACpF,QAAQ;YACR,UAAU,CAAC,QAAQ;;;;;MAMzB;QACE,IAAI;QACJ,MAAM;QACN,aAAa;QACb,WAAW,EAAE,OAAO,KAAK,QAAQ,IAAG;QACpC,QAAQ,QAAA;QACR,SAAS;UACP,MAAM;YACJ,MAAK,GAAA,QAAA,YAAW,SAAS,wDAAwD;YACjF,QAAQ;;UAEV,QAAQ;YACN,MAAK,GAAA,QAAA,YAAW,SAAS,+DAA+D;YACxF,QAAQ;YACR,aAAa;YACb,OAAO,6BAAA;YACP,UAAU,CAAC,QAAQ;;UAErB,UAAU;YACR,MAAK,GAAA,QAAA,YAAW,SAAS,2DAA2D;YACpF,QAAQ;YACR,UAAU,CAAC,QAAQ;;;;MAIzB;QACE,IAAI;QACJ,MAAM;QACN,aAAa;QACb,WAAW,EAAE,OAAO,KAAK,QAAQ,IAAG;QACpC,QAAQ,QAAA;QACR,SAAS;UACP,MAAM;YACJ,MAAK,GAAA,QAAA,YAAW,SAAS,wDAAwD;YACjF,QAAQ;;UAEV,QAAQ;YACN,MAAK,GAAA,QAAA,YAAW,SAAS,+DAA+D;YACxF,QAAQ;YACR,aAAa;YACb,OAAO,6BAAA;YACP,UAAU,CAAC,QAAQ;;UAErB,UAAU;YACR,MAAK,GAAA,QAAA,YAAW,SAAS,2DAA2D;YACpF,QAAQ;YACR,UAAU,CAAC,QAAQ;;;;MAIzB;QACE,IAAI;QACJ,MAAM;QACN,aAAa;QACb,WAAW,EAAE,OAAO,KAAK,QAAQ,IAAG;QACpC,QAAQ,QAAA;QACR,SAAS;UACP,MAAM;YACJ,MAAK,GAAA,QAAA,YAAW,SAAS,wDAAwD;YACjF,QAAQ;;UAEV,QAAQ;YACN,MAAK,GAAA,QAAA,YAAW,SAAS,+DAA+D;YACxF,QAAQ;YACR,aAAa;YACb,OAAO,6BAAA;YACP,UAAU,CAAC,QAAQ;;UAErB,UAAU;YACR,MAAK,GAAA,QAAA,YAAW,SAAS,2DAA2D;YACpF,QAAQ;YACR,UAAU,CAAC,QAAQ;;;;;;;;;;;;;ACnI3B,YAAA,UAAA;AAcA,YAAA,iBAAA;AAmCA,YAAA,uBAAA;AA2CA,YAAA,qBAAA;AApHA,QAAA,YAAA;AAwBA,aAAgB,QAAQ,GAAS;AAC/B,aAAO,KAAK,IAAI,KAAK,IAAI,CAAC,CAAC;IAC7B;AAYA,aAAgB,eACd,QACA,QACA,eACA,OACA,OAAa;AAEb,YAAM,WAAW,QAAQ;AACzB,YAAM,UAAU,IAAI,aAAa,QAAQ;AAIzC,eAAS,KAAK,GAAG,KAAK,UAAU,MAAM;AACpC,YAAI,MAAM;AACV,iBAAS,IAAI,GAAG,IAAI,eAAe,KAAK;AACtC,kBAAQ,OAAO,CAAC,KAAK,MAAM,OAAO,IAAI,WAAW,EAAE,KAAK;QAC1D;AACA,gBAAQ,EAAE,IAAI,QAAQ,GAAG;MAC3B;AAEA,aAAO;IACT;AAcA,aAAgB,qBACd,SACA,OACA,OACA,MACA,eACA,WAAiB;AAEjB,YAAM,SAAS,KAAK,IAAI,GAAG,KAAK,MAAM,KAAK,IAAI,SAAS,CAAC;AACzD,YAAM,SAAS,KAAK,IAAI,GAAG,KAAK,MAAM,KAAK,IAAI,SAAS,CAAC;AACzD,YAAM,SAAS,KAAK,IAAI,OAAO,KAAK,MAAM,KAAK,IAAI,KAAK,KAAK,SAAS,CAAC;AACvE,YAAM,SAAS,KAAK,IAAI,OAAO,KAAK,MAAM,KAAK,IAAI,KAAK,KAAK,SAAS,CAAC;AAEvE,YAAM,QAAQ,KAAK,IAAI,GAAG,SAAS,MAAM;AACzC,YAAM,QAAQ,KAAK,IAAI,GAAG,SAAS,MAAM;AAEzC,YAAM,OAAO,IAAI,WAAW,QAAQ,KAAK;AAEzC,eAAS,MAAM,GAAG,MAAM,OAAO,OAAO;AACpC,cAAM,SAAS,SAAS;AACxB,iBAAS,MAAM,GAAG,MAAM,OAAO,OAAO;AACpC,gBAAM,SAAS,SAAS;AACxB,gBAAM,SAAS,SAAS,QAAQ;AAChC,eAAK,MAAM,QAAQ,GAAG,KAAK,QAAQ,MAAM,KAAK,KAAK,gBAAgB,MAAM;QAC3E;MACF;AAEA,aAAO,EAAE,MAAM,OAAO,OAAO,QAAQ,MAAK;IAC5C;AAeA,aAAgB,mBACd,WACA,SAAkC;AAElC,YAAM,EACJ,iBACA,aACA,YACA,UACA,eACA,YACA,UAAS,IACP;AAEJ,YAAM,EACJ,YACA,cACA,QACA,OACA,MACA,MACA,eACA,gBACA,gBAAgB,IAAG,IACjB;AAOJ,YAAM,gBAAgB;AACtB,YAAM,YAAY,aAAa;AAS/B,YAAM,aAA0B,CAAA;AAEhC,eAAS,IAAI,GAAG,IAAI,UAAU,KAAK;AACjC,cAAM,KAAK,gBAAgB,IAAI,WAAW,CAAC,KAAK;AAChD,cAAM,KAAK,gBAAgB,IAAI,WAAW,CAAC,KAAK;AAChD,cAAM,IAAI,gBAAgB,IAAI,WAAW,CAAC,KAAK;AAC/C,cAAM,IAAI,gBAAgB,IAAI,WAAW,CAAC,KAAK;AAE/C,YAAI,YAAY;AAChB,YAAI,YAAY;AAEhB,iBAAS,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,gBAAM,QAAQ,iBAAiB,IAAI,KAAK,WAAW,CAAC,KAAK;AACzD,cAAI,QAAQ,WAAW;AACrB,wBAAY;AACZ,wBAAY;UACd;QACF;AAEA,YAAI,YAAY;AAAY;AAE5B,cAAM,OAAoB;UACxB,GAAG,KAAK,IAAI;UACZ,GAAG,KAAK,IAAI;UACZ;UACA;;AAIF,cAAM,SAAS,IAAI,aAAa,aAAa;AAC7C,iBAAS,IAAI,GAAG,IAAI,eAAe,KAAK;AACtC,iBAAO,CAAC,IAAI,iBAAiB,IAAI,aAAa,KAAK,WAAW,CAAC,KAAK;QACtE;AAEA,mBAAW,KAAK,EAAE,MAAM,OAAO,WAAW,UAAU,WAAW,OAAM,CAAE;MACzE;AAEA,UAAI,WAAW,WAAW;AAAG,eAAO,CAAA;AAEpC,YAAM,eAAc,GAAA,UAAA,KAAI,YAAY,YAAY;AAEhD,aAAO,YAAY,IAAI,CAAC,QAAO;AAC7B,cAAM,EAAE,MAAM,OAAO,UAAU,OAAM,IAAK,WAAW,GAAG;AACxD,cAAM,QAAQ,OAAO,QAAQ,KAAK,OAAO,QAAQ;AAGjD,cAAM,IAAI,KAAK,IAAI,GAAG,KAAK,IAAI,gBAAgB,KAAK,IAAI,QAAQ,KAAK,CAAC;AACtE,cAAM,IAAI,KAAK,IAAI,GAAG,KAAK,IAAI,iBAAiB,KAAK,IAAI,QAAQ,KAAK,CAAC;AACvE,cAAM,KAAK,KAAK,IAAI,GAAG,KAAK,IAAI,gBAAgB,KAAK,IAAI,KAAK,IAAI,QAAQ,KAAK,CAAC;AAChF,cAAM,KAAK,KAAK,IAAI,GAAG,KAAK,IAAI,iBAAiB,KAAK,IAAI,KAAK,IAAI,QAAQ,KAAK,CAAC;AACjF,cAAM,YAAyB,EAAE,GAAG,GAAG,GAAG,KAAK,GAAG,GAAG,KAAK,EAAC;AAG3D,cAAM,UAAU,eAAe,QAAQ,aAAa,eAAe,YAAY,SAAS;AACxF,cAAM,EAAE,MAAM,UAAU,OAAO,IAAI,QAAQ,GAAE,IAAK,qBAChD,SACA,YACA,WACA,MACA,eACA,SAAS;AAGX,eAAO;UACL,OAAO;UACP,eAAe;UACf;UACA,MAAM;UACN,MAAM;UACN,WAAW;UACX,YAAY;;MAEhB,CAAC;IACH;;;","names":[]}