tempest-react-sdk 0.23.0 → 0.24.0

package/dist/vision/postprocess/segmentation.js

@@ -1,109 +1,103 @@
-import { decodeYoloAnchors as D } from "./detection.js";
-import { Mask as U, BoundingBox as ee } from "../types.js";
-function ne(t, n, l, s, e) {
-  let o = s;
-  if (o.length === 4) {
-    if (o[0] !== 1)
-      throw new Error(`decodeYoloSeg: expected batch size 1 in prototypes, got ${o[0]}.`);
-    o = [o[1], o[2], o[3]];
-  }
-  if (o.length !== 3)
-    throw new Error(
-      `decodeYoloSeg: expected 3-D prototypes after batch removal, got dims=${JSON.stringify(s)}.`
-    );
-  const m = o[0], i = o[1], r = o[2], u = n.length === 3 ? n[1] : n[0], h = n.length === 3 ? n[2] : n[1];
-  if (u === void 0 || h === void 0)
-    throw new Error(
-      `decodeYoloSeg: cannot read channels/numAnchors from dims=${JSON.stringify(n)}.`
-    );
-  const g = 4 + e.numClasses + m;
-  if (u !== g)
-    throw new Error(
-      `decodeYoloSeg: channels=${u} does not match 4 + numClasses(${e.numClasses}) + numMaskCoefs(${m}) = ${g}.`
-    );
-  if (l.length !== m * i * r)
-    throw new Error(
-      `decodeYoloSeg: prototype length ${l.length} does not match dims=${JSON.stringify(s)}.`
-    );
-  const c = D(t, n, {
-    numClasses: e.numClasses,
-    originalWidth: e.originalWidth,
-    originalHeight: e.originalHeight,
-    padLeft: e.padLeft,
-    padTop: e.padTop,
-    scale: e.scale,
-    confThreshold: e.confThreshold,
-    iouThreshold: e.iouThreshold,
-    maxDetections: e.maxDetections
-  });
-  if (c.anchorIndices.length === 0) return [];
-  const x = e.maskThreshold ?? 0.5, k = r / e.inputWidth, f = i / e.inputHeight, C = i * r, y = (4 + e.numClasses) * h, M = [];
-  for (let a = 0; a < c.anchorIndices.length; a++) {
-    const I = c.anchorIndices[a], S = c.boxesXyxy[a * 4], Y = c.boxesXyxy[a * 4 + 1], T = c.boxesXyxy[a * 4 + 2], X = c.boxesXyxy[a * 4 + 3], B = new ee(S, Y, T, X), L = c.classIds[a], p = c.confidences[a], w = Math.max(0, Math.trunc(T) - Math.trunc(S)), b = Math.max(0, Math.trunc(X) - Math.trunc(Y));
-    if (w === 0 || b === 0) {
-      M.push({ bbox: B, classId: L, confidence: p, mask: new U(new Uint8Array(0), 0, 0) });
-      continue;
-    }
-    const q = S * e.scale + e.padLeft, A = Y * e.scale + e.padTop, G = T * e.scale + e.padLeft, K = X * e.scale + e.padTop, z = Math.max(0, Math.floor(q * k)), J = Math.max(0, Math.floor(A * f)), V = Math.min(r, Math.ceil(G * k)), F = Math.min(i, Math.ceil(K * f));
-    if (V <= z || F <= J) {
-      M.push({
-        bbox: B,
-        classId: L,
-        confidence: p,
-        mask: new U(new Uint8Array(w * b), w, b)
-      });
-      continue;
-    }
-    const v = V - z, N = F - J, j = new Float32Array(v * N);
-    for (let d = 0; d < N; d++) {
-      const R = J + d;
-      for (let $ = 0; $ < v; $++) {
-        const Z = z + $;
-        let P = 0;
-        for (let E = 0; E < m; E++) {
-          const H = t[y + E * h + I], W = l[E * C + R * r + Z];
-          P += H * W;
-        }
-        j[d * v + $] = oe(P);
-      }
-    }
-    const Q = te(j, v, N, w, b), O = new Uint8Array(w * b);
-    for (let d = 0; d < O.length; d++)
-      O[d] = Q[d] >= x ? 255 : 0;
-    M.push({ bbox: B, classId: L, confidence: p, mask: new U(O, w, b) });
-  }
-  return M;
+import { BoundingBox as e, Mask as t } from "../types.js";
+import { decodeYoloAnchors as n } from "./detection.js";
+//#region src/vision/postprocess/segmentation.ts
+function r(r, a, o, c, l) {
+	let u = c;
+	if (u.length === 4) {
+		if (u[0] !== 1) throw Error(`decodeYoloSeg: expected batch size 1 in prototypes, got ${u[0]}.`);
+		u = [
+			u[1],
+			u[2],
+			u[3]
+		];
+	}
+	if (u.length !== 3) throw Error(`decodeYoloSeg: expected 3-D prototypes after batch removal, got dims=${JSON.stringify(c)}.`);
+	let d = u[0], f = u[1], p = u[2], m = a.length === 3 ? a[1] : a[0], h = a.length === 3 ? a[2] : a[1];
+	if (m === void 0 || h === void 0) throw Error(`decodeYoloSeg: cannot read channels/numAnchors from dims=${JSON.stringify(a)}.`);
+	let g = 4 + l.numClasses + d;
+	if (m !== g) throw Error(`decodeYoloSeg: channels=${m} does not match 4 + numClasses(${l.numClasses}) + numMaskCoefs(${d}) = ${g}.`);
+	if (o.length !== d * f * p) throw Error(`decodeYoloSeg: prototype length ${o.length} does not match dims=${JSON.stringify(c)}.`);
+	let _ = n(r, a, {
+		numClasses: l.numClasses,
+		originalWidth: l.originalWidth,
+		originalHeight: l.originalHeight,
+		padLeft: l.padLeft,
+		padTop: l.padTop,
+		scale: l.scale,
+		confThreshold: l.confThreshold,
+		iouThreshold: l.iouThreshold,
+		maxDetections: l.maxDetections
+	});
+	if (_.anchorIndices.length === 0) return [];
+	let v = l.maskThreshold ?? .5, y = p / l.inputWidth, b = f / l.inputHeight, x = f * p, S = (4 + l.numClasses) * h, C = [];
+	for (let n = 0; n < _.anchorIndices.length; n++) {
+		let a = _.anchorIndices[n], c = _.boxesXyxy[n * 4], u = _.boxesXyxy[n * 4 + 1], m = _.boxesXyxy[n * 4 + 2], g = _.boxesXyxy[n * 4 + 3], w = new e(c, u, m, g), T = _.classIds[n], E = _.confidences[n], D = Math.max(0, Math.trunc(m) - Math.trunc(c)), O = Math.max(0, Math.trunc(g) - Math.trunc(u));
+		if (D === 0 || O === 0) {
+			C.push({
+				bbox: w,
+				classId: T,
+				confidence: E,
+				mask: new t(/* @__PURE__ */ new Uint8Array(), 0, 0)
+			});
+			continue;
+		}
+		let k = c * l.scale + l.padLeft, A = u * l.scale + l.padTop, j = m * l.scale + l.padLeft, M = g * l.scale + l.padTop, N = Math.max(0, Math.floor(k * y)), P = Math.max(0, Math.floor(A * b)), F = Math.min(p, Math.ceil(j * y)), I = Math.min(f, Math.ceil(M * b));
+		if (F <= N || I <= P) {
+			C.push({
+				bbox: w,
+				classId: T,
+				confidence: E,
+				mask: new t(new Uint8Array(D * O), D, O)
+			});
+			continue;
+		}
+		let L = F - N, R = I - P, z = new Float32Array(L * R);
+		for (let e = 0; e < R; e++) {
+			let t = P + e;
+			for (let n = 0; n < L; n++) {
+				let s = N + n, c = 0;
+				for (let e = 0; e < d; e++) {
+					let n = r[S + e * h + a], i = o[e * x + t * p + s];
+					c += n * i;
+				}
+				z[e * L + n] = i(c);
+			}
+		}
+		let B = s(z, L, R, D, O), V = new Uint8Array(D * O);
+		for (let e = 0; e < V.length; e++) V[e] = B[e] >= v ? 255 : 0;
+		C.push({
+			bbox: w,
+			classId: T,
+			confidence: E,
+			mask: new t(V, D, O)
+		});
+	}
+	return C;
 }
-function oe(t) {
-  if (t >= 0)
-    return 1 / (1 + Math.exp(-t));
-  const n = Math.exp(t);
-  return n / (1 + n);
+function i(e) {
+	if (e >= 0) return 1 / (1 + Math.exp(-e));
+	let t = Math.exp(e);
+	return t / (1 + t);
 }
-let _ = !1;
-function ae(t, n, l, s, e) {
-  return _ || (_ = !0, console.warn(
-    "[@ort-vision-sdk/web] decodeYoloV8Seg is deprecated since 0.2.0; use decodeYoloSeg. The alias will be removed in 0.3.0."
-  )), ne(t, n, l, s, e);
+var a = !1;
+function o(e, t, n, i, o) {
+	return a || (a = !0, console.warn("[@ort-vision-sdk/web] decodeYoloV8Seg is deprecated since 0.2.0; use decodeYoloSeg. The alias will be removed in 0.3.0.")), r(e, t, n, i, o);
 }
-function te(t, n, l, s, e) {
-  const o = new Float32Array(s * e);
-  if (s === 0 || e === 0 || n === 0 || l === 0)
-    return o;
-  if (s === n && e === l)
-    return o.set(t), o;
-  const m = n / s, i = l / e;
-  for (let r = 0; r < e; r++) {
-    const u = (r + 0.5) * i - 0.5, h = Math.max(0, Math.floor(u)), g = Math.min(l - 1, h + 1), c = Math.max(0, Math.min(1, u - h));
-    for (let x = 0; x < s; x++) {
-      const k = (x + 0.5) * m - 0.5, f = Math.max(0, Math.floor(k)), C = Math.min(n - 1, f + 1), y = Math.max(0, Math.min(1, k - f)), M = t[h * n + f], a = t[h * n + C], I = t[g * n + f], S = t[g * n + C], Y = M * (1 - y) + a * y, T = I * (1 - y) + S * y;
-      o[r * s + x] = Y * (1 - c) + T * c;
-    }
-  }
-  return o;
+function s(e, t, n, r, i) {
+	let a = new Float32Array(r * i);
+	if (r === 0 || i === 0 || t === 0 || n === 0) return a;
+	if (r === t && i === n) return a.set(e), a;
+	let o = t / r, s = n / i;
+	for (let c = 0; c < i; c++) {
+		let i = (c + .5) * s - .5, l = Math.max(0, Math.floor(i)), u = Math.min(n - 1, l + 1), d = Math.max(0, Math.min(1, i - l));
+		for (let n = 0; n < r; n++) {
+			let i = (n + .5) * o - .5, s = Math.max(0, Math.floor(i)), f = Math.min(t - 1, s + 1), p = Math.max(0, Math.min(1, i - s)), m = e[l * t + s], h = e[l * t + f], g = e[u * t + s], _ = e[u * t + f], v = m * (1 - p) + h * p, y = g * (1 - p) + _ * p;
+			a[c * r + n] = v * (1 - d) + y * d;
+		}
+	}
+	return a;
 }
-export {
-  ne as decodeYoloSeg,
-  ae as decodeYoloV8Seg
-};
-//# sourceMappingURL=segmentation.js.map
+//#endregion
+export { r as decodeYoloSeg, o as decodeYoloV8Seg };
+
+//# sourceMappingURL=segmentation.js.map

package/dist/vision/postprocess/segmentation.js.map

@@ -1 +1 @@
-{"version":3,"file":"segmentation.js","sources":["../../../src/vision/postprocess/segmentation.ts"],"sourcesContent":["/**\n * Segmentation head postprocessing: YOLO instance-segmentation decoding.\n *\n * Compatible with YOLOv8-seg / YOLOv11-seg exports (and any later seg head\n * sharing the layout). The model produces two output tensors:\n *\n * - `output0` of shape `(1, 4 + numClasses + numMaskCoefs, numAnchors)` — the\n *   same per-anchor predictions as plain YOLO detection plus an extra\n *   `numMaskCoefs` (typically 32) channels of mask coefficients.\n * - `output1` of shape `(1, numMaskCoefs, maskH, maskW)` — a set of \"prototype\"\n *   masks shared across all anchors.\n *\n * The per-anchor decode (xywh→xyxy, undo letterbox, per-class NMS, sort & cap)\n * is delegated to {@link decodeYoloAnchors}; this module only handles the\n * mask-specific work: matmul against prototypes, sigmoid, bilinear resize and\n * thresholding.\n */\n\nimport { decodeYoloAnchors } from \"./detection\";\nimport { BoundingBox, Mask } from \"../types\";\n\nexport interface DecodeYoloSegOptions {\n    readonly numClasses: number;\n    /** Model input `[width, height]` (post-letterbox). */\n    readonly inputWidth: number;\n    readonly inputHeight: number;\n    /** Original image `[width, height]`. */\n    readonly originalWidth: number;\n    readonly originalHeight: number;\n    /** Letterbox horizontal padding in input-tensor pixels. */\n    readonly padLeft: number;\n    /** Letterbox vertical padding in input-tensor pixels. */\n    readonly padTop: number;\n    /** Letterbox scale factor. */\n    readonly scale: number;\n    readonly confThreshold: number;\n    readonly iouThreshold: number;\n    readonly maxDetections: number;\n    /** Probability cutoff applied to the soft mask. Defaults to `0.5`. */\n    readonly maskThreshold?: number;\n}\n\nexport interface DecodedSegmentation {\n    readonly bbox: BoundingBox;\n    readonly classId: number;\n    readonly confidence: number;\n    /** Binary mask cropped to `bbox`. Width/height match `bbox.asIntXyxy()` extents. */\n    readonly mask: Mask;\n}\n\n/**\n * Decode YOLO segmentation raw outputs into a list of segmented instances.\n *\n * Compatible with YOLOv8-seg / YOLOv11-seg.\n *\n * @param perAnchorData Flat `output0`, length `(4 + numClasses + numMaskCoefs) * numAnchors`.\n * @param perAnchorDims Dims as reported by ORT, e.g. `[1, 116, 8400]`.\n * @param prototypeData Flat `output1`, length `numMaskCoefs * maskH * maskW`.\n * @param prototypeDims Dims as reported by ORT, e.g. `[1, 32, 160, 160]`.\n */\nexport function decodeYoloSeg(\n    perAnchorData: Float32Array,\n    perAnchorDims: readonly number[],\n    prototypeData: Float32Array,\n    prototypeDims: readonly number[],\n    options: DecodeYoloSegOptions,\n): DecodedSegmentation[] {\n    // Strip batch dim from prototypes; perAnchor batch is validated by the helper.\n    let pDims = prototypeDims;\n    if (pDims.length === 4) {\n        if (pDims[0] !== 1) {\n            throw new Error(`decodeYoloSeg: expected batch size 1 in prototypes, got ${pDims[0]}.`);\n        }\n        pDims = [pDims[1] as number, pDims[2] as number, pDims[3] as number];\n    }\n    if (pDims.length !== 3) {\n        throw new Error(\n            `decodeYoloSeg: expected 3-D prototypes after batch removal, got dims=${JSON.stringify(prototypeDims)}.`,\n        );\n    }\n    const numMaskCoefs = pDims[0] as number;\n    const maskH = pDims[1] as number;\n    const maskW = pDims[2] as number;\n\n    const channels = perAnchorDims.length === 3 ? perAnchorDims[1] : perAnchorDims[0];\n    const numAnchors = perAnchorDims.length === 3 ? perAnchorDims[2] : perAnchorDims[1];\n    if (channels === undefined || numAnchors === undefined) {\n        throw new Error(\n            `decodeYoloSeg: cannot read channels/numAnchors from dims=${JSON.stringify(perAnchorDims)}.`,\n        );\n    }\n\n    const expectedChannels = 4 + options.numClasses + numMaskCoefs;\n    if (channels !== expectedChannels) {\n        throw new Error(\n            `decodeYoloSeg: channels=${channels} does not match 4 + numClasses(${options.numClasses}) + numMaskCoefs(${numMaskCoefs}) = ${expectedChannels}.`,\n        );\n    }\n    if (prototypeData.length !== numMaskCoefs * maskH * maskW) {\n        throw new Error(\n            `decodeYoloSeg: prototype length ${prototypeData.length} does not match dims=${JSON.stringify(prototypeDims)}.`,\n        );\n    }\n\n    const decoded = decodeYoloAnchors(perAnchorData, perAnchorDims, {\n        numClasses: options.numClasses,\n        originalWidth: options.originalWidth,\n        originalHeight: options.originalHeight,\n        padLeft: options.padLeft,\n        padTop: options.padTop,\n        scale: options.scale,\n        confThreshold: options.confThreshold,\n        iouThreshold: options.iouThreshold,\n        maxDetections: options.maxDetections,\n    });\n\n    if (decoded.anchorIndices.length === 0) return [];\n\n    const maskThreshold = options.maskThreshold ?? 0.5;\n    const scaleX = maskW / options.inputWidth;\n    const scaleY = maskH / options.inputHeight;\n    const protoPlane = maskH * maskW;\n    const coefBase = (4 + options.numClasses) * numAnchors;\n\n    const results: DecodedSegmentation[] = [];\n\n    for (let i = 0; i < decoded.anchorIndices.length; i++) {\n        const a = decoded.anchorIndices[i] as number;\n        const x1 = decoded.boxesXyxy[i * 4] as number;\n        const y1 = decoded.boxesXyxy[i * 4 + 1] as number;\n        const x2 = decoded.boxesXyxy[i * 4 + 2] as number;\n        const y2 = decoded.boxesXyxy[i * 4 + 3] as number;\n        const bbox = new BoundingBox(x1, y1, x2, y2);\n        const classId = decoded.classIds[i] as number;\n        const confidence = decoded.confidences[i] as number;\n\n        const bboxW = Math.max(0, Math.trunc(x2) - Math.trunc(x1));\n        const bboxH = Math.max(0, Math.trunc(y2) - Math.trunc(y1));\n\n        if (bboxW === 0 || bboxH === 0) {\n            results.push({ bbox, classId, confidence, mask: new Mask(new Uint8Array(0), 0, 0) });\n            continue;\n        }\n\n        // Bbox in input-tensor coords, then in low-res mask coords.\n        const ibx1 = x1 * options.scale + options.padLeft;\n        const iby1 = y1 * options.scale + options.padTop;\n        const ibx2 = x2 * options.scale + options.padLeft;\n        const iby2 = y2 * options.scale + options.padTop;\n\n        const mbx1 = Math.max(0, Math.floor(ibx1 * scaleX));\n        const mby1 = Math.max(0, Math.floor(iby1 * scaleY));\n        const mbx2 = Math.min(maskW, Math.ceil(ibx2 * scaleX));\n        const mby2 = Math.min(maskH, Math.ceil(iby2 * scaleY));\n\n        if (mbx2 <= mbx1 || mby2 <= mby1) {\n            results.push({\n                bbox,\n                classId,\n                confidence,\n                mask: new Mask(new Uint8Array(bboxW * bboxH), bboxW, bboxH),\n            });\n            continue;\n        }\n\n        // Compute soft mask only within the prototype region under this bbox.\n        const cropW = mbx2 - mbx1;\n        const cropH = mby2 - mby1;\n        const softCrop = new Float32Array(cropW * cropH);\n        for (let y = 0; y < cropH; y++) {\n            const py = mby1 + y;\n            for (let x = 0; x < cropW; x++) {\n                const px = mbx1 + x;\n                let sum = 0;\n                for (let kk = 0; kk < numMaskCoefs; kk++) {\n                    const coef = perAnchorData[coefBase + kk * numAnchors + a];\n                    const proto = prototypeData[kk * protoPlane + py * maskW + px];\n                    sum += coef * proto;\n                }\n                softCrop[y * cropW + x] = sigmoid(sum);\n            }\n        }\n\n        const resized = resizeBilinear(softCrop, cropW, cropH, bboxW, bboxH);\n        const binary = new Uint8Array(bboxW * bboxH);\n        for (let j = 0; j < binary.length; j++) {\n            binary[j] = resized[j] >= maskThreshold ? 255 : 0;\n        }\n\n        results.push({ bbox, classId, confidence, mask: new Mask(binary, bboxW, bboxH) });\n    }\n\n    return results;\n}\n\nfunction sigmoid(x: number): number {\n    if (x >= 0) {\n        return 1 / (1 + Math.exp(-x));\n    }\n    const e = Math.exp(x);\n    return e / (1 + e);\n}\n\nlet _warnedDecodeYoloV8Seg = false;\n\n/** @deprecated since 0.2.0 — use {@link decodeYoloSeg}. Will be removed in 0.3.0. */\nexport function decodeYoloV8Seg(\n    perAnchorData: Float32Array,\n    perAnchorDims: readonly number[],\n    prototypeData: Float32Array,\n    prototypeDims: readonly number[],\n    options: DecodeYoloSegOptions,\n): DecodedSegmentation[] {\n    if (!_warnedDecodeYoloV8Seg) {\n        _warnedDecodeYoloV8Seg = true;\n        console.warn(\n            \"[@ort-vision-sdk/web] decodeYoloV8Seg is deprecated since 0.2.0; use decodeYoloSeg. \" +\n                \"The alias will be removed in 0.3.0.\",\n        );\n    }\n    return decodeYoloSeg(perAnchorData, perAnchorDims, prototypeData, prototypeDims, options);\n}\n\n/** @deprecated since 0.2.0 — use {@link DecodeYoloSegOptions}. */\nexport type DecodeYoloV8SegOptions = DecodeYoloSegOptions;\n\nfunction resizeBilinear(\n    src: Float32Array,\n    srcWidth: number,\n    srcHeight: number,\n    targetWidth: number,\n    targetHeight: number,\n): Float32Array {\n    const out = new Float32Array(targetWidth * targetHeight);\n    if (targetWidth === 0 || targetHeight === 0 || srcWidth === 0 || srcHeight === 0) {\n        return out;\n    }\n    if (targetWidth === srcWidth && targetHeight === srcHeight) {\n        out.set(src);\n        return out;\n    }\n    const sx = srcWidth / targetWidth;\n    const sy = srcHeight / targetHeight;\n    for (let y = 0; y < targetHeight; y++) {\n        const yy = (y + 0.5) * sy - 0.5;\n        const y0 = Math.max(0, Math.floor(yy));\n        const y1 = Math.min(srcHeight - 1, y0 + 1);\n        const wy = Math.max(0, Math.min(1, yy - y0));\n        for (let x = 0; x < targetWidth; x++) {\n            const xx = (x + 0.5) * sx - 0.5;\n            const x0 = Math.max(0, Math.floor(xx));\n            const x1 = Math.min(srcWidth - 1, x0 + 1);\n            const wx = Math.max(0, Math.min(1, xx - x0));\n            const v00 = src[y0 * srcWidth + x0];\n            const v01 = src[y0 * srcWidth + x1];\n            const v10 = src[y1 * srcWidth + x0];\n            const v11 = src[y1 * srcWidth + x1];\n            const top = v00 * (1 - wx) + v01 * wx;\n            const bot = v10 * (1 - wx) + v11 * wx;\n            out[y * targetWidth + x] = top * (1 - wy) + bot * wy;\n        }\n    }\n    return out;\n}\n"],"names":["decodeYoloSeg","perAnchorData","perAnchorDims","prototypeData","prototypeDims","options","pDims","numMaskCoefs","maskH","maskW","channels","numAnchors","expectedChannels","decoded","decodeYoloAnchors","maskThreshold","scaleX","scaleY","protoPlane","coefBase","results","i","a","x1","y1","x2","y2","bbox","BoundingBox","classId","confidence","bboxW","bboxH","Mask","ibx1","iby1","ibx2","iby2","mbx1","mby1","mbx2","mby2","cropW","cropH","softCrop","y","py","x","px","sum","kk","coef","proto","sigmoid","resized","resizeBilinear","binary","j","e","_warnedDecodeYoloV8Seg","decodeYoloV8Seg","src","srcWidth","srcHeight","targetWidth","targetHeight","out","sx","sy","yy","y0","wy","xx","x0","wx","v00","v01","v10","v11","top","bot"],"mappings":";;AA4DO,SAASA,GACZC,GACAC,GACAC,GACAC,GACAC,GACqB;AAErB,MAAIC,IAAQF;AACZ,MAAIE,EAAM,WAAW,GAAG;AACpB,QAAIA,EAAM,CAAC,MAAM;AACb,YAAM,IAAI,MAAM,2DAA2DA,EAAM,CAAC,CAAC,GAAG;AAE1F,IAAAA,IAAQ,CAACA,EAAM,CAAC,GAAaA,EAAM,CAAC,GAAaA,EAAM,CAAC,CAAW;AAAA,EACvE;AACA,MAAIA,EAAM,WAAW;AACjB,UAAM,IAAI;AAAA,MACN,wEAAwE,KAAK,UAAUF,CAAa,CAAC;AAAA,IAAA;AAG7G,QAAMG,IAAeD,EAAM,CAAC,GACtBE,IAAQF,EAAM,CAAC,GACfG,IAAQH,EAAM,CAAC,GAEfI,IAAWR,EAAc,WAAW,IAAIA,EAAc,CAAC,IAAIA,EAAc,CAAC,GAC1ES,IAAaT,EAAc,WAAW,IAAIA,EAAc,CAAC,IAAIA,EAAc,CAAC;AAClF,MAAIQ,MAAa,UAAaC,MAAe;AACzC,UAAM,IAAI;AAAA,MACN,4DAA4D,KAAK,UAAUT,CAAa,CAAC;AAAA,IAAA;AAIjG,QAAMU,IAAmB,IAAIP,EAAQ,aAAaE;AAClD,MAAIG,MAAaE;AACb,UAAM,IAAI;AAAA,MACN,2BAA2BF,CAAQ,kCAAkCL,EAAQ,UAAU,oBAAoBE,CAAY,OAAOK,CAAgB;AAAA,IAAA;AAGtJ,MAAIT,EAAc,WAAWI,IAAeC,IAAQC;AAChD,UAAM,IAAI;AAAA,MACN,mCAAmCN,EAAc,MAAM,wBAAwB,KAAK,UAAUC,CAAa,CAAC;AAAA,IAAA;AAIpH,QAAMS,IAAUC,EAAkBb,GAAeC,GAAe;AAAA,IAC5D,YAAYG,EAAQ;AAAA,IACpB,eAAeA,EAAQ;AAAA,IACvB,gBAAgBA,EAAQ;AAAA,IACxB,SAASA,EAAQ;AAAA,IACjB,QAAQA,EAAQ;AAAA,IAChB,OAAOA,EAAQ;AAAA,IACf,eAAeA,EAAQ;AAAA,IACvB,cAAcA,EAAQ;AAAA,IACtB,eAAeA,EAAQ;AAAA,EAAA,CAC1B;AAED,MAAIQ,EAAQ,cAAc,WAAW,UAAU,CAAA;AAE/C,QAAME,IAAgBV,EAAQ,iBAAiB,KACzCW,IAASP,IAAQJ,EAAQ,YACzBY,IAAST,IAAQH,EAAQ,aACzBa,IAAaV,IAAQC,GACrBU,KAAY,IAAId,EAAQ,cAAcM,GAEtCS,IAAiC,CAAA;AAEvC,WAASC,IAAI,GAAGA,IAAIR,EAAQ,cAAc,QAAQQ,KAAK;AACnD,UAAMC,IAAIT,EAAQ,cAAcQ,CAAC,GAC3BE,IAAKV,EAAQ,UAAUQ,IAAI,CAAC,GAC5BG,IAAKX,EAAQ,UAAUQ,IAAI,IAAI,CAAC,GAChCI,IAAKZ,EAAQ,UAAUQ,IAAI,IAAI,CAAC,GAChCK,IAAKb,EAAQ,UAAUQ,IAAI,IAAI,CAAC,GAChCM,IAAO,IAAIC,GAAYL,GAAIC,GAAIC,GAAIC,CAAE,GACrCG,IAAUhB,EAAQ,SAASQ,CAAC,GAC5BS,IAAajB,EAAQ,YAAYQ,CAAC,GAElCU,IAAQ,KAAK,IAAI,GAAG,KAAK,MAAMN,CAAE,IAAI,KAAK,MAAMF,CAAE,CAAC,GACnDS,IAAQ,KAAK,IAAI,GAAG,KAAK,MAAMN,CAAE,IAAI,KAAK,MAAMF,CAAE,CAAC;AAEzD,QAAIO,MAAU,KAAKC,MAAU,GAAG;AAC5B,MAAAZ,EAAQ,KAAK,EAAE,MAAAO,GAAM,SAAAE,GAAS,YAAAC,GAAY,MAAM,IAAIG,EAAK,IAAI,WAAW,CAAC,GAAG,GAAG,CAAC,GAAG;AACnF;AAAA,IACJ;AAGA,UAAMC,IAAOX,IAAKlB,EAAQ,QAAQA,EAAQ,SACpC8B,IAAOX,IAAKnB,EAAQ,QAAQA,EAAQ,QACpC+B,IAAOX,IAAKpB,EAAQ,QAAQA,EAAQ,SACpCgC,IAAOX,IAAKrB,EAAQ,QAAQA,EAAQ,QAEpCiC,IAAO,KAAK,IAAI,GAAG,KAAK,MAAMJ,IAAOlB,CAAM,CAAC,GAC5CuB,IAAO,KAAK,IAAI,GAAG,KAAK,MAAMJ,IAAOlB,CAAM,CAAC,GAC5CuB,IAAO,KAAK,IAAI/B,GAAO,KAAK,KAAK2B,IAAOpB,CAAM,CAAC,GAC/CyB,IAAO,KAAK,IAAIjC,GAAO,KAAK,KAAK6B,IAAOpB,CAAM,CAAC;AAErD,QAAIuB,KAAQF,KAAQG,KAAQF,GAAM;AAC9B,MAAAnB,EAAQ,KAAK;AAAA,QACT,MAAAO;AAAA,QACA,SAAAE;AAAA,QACA,YAAAC;AAAA,QACA,MAAM,IAAIG,EAAK,IAAI,WAAWF,IAAQC,CAAK,GAAGD,GAAOC,CAAK;AAAA,MAAA,CAC7D;AACD;AAAA,IACJ;AAGA,UAAMU,IAAQF,IAAOF,GACfK,IAAQF,IAAOF,GACfK,IAAW,IAAI,aAAaF,IAAQC,CAAK;AAC/C,aAASE,IAAI,GAAGA,IAAIF,GAAOE,KAAK;AAC5B,YAAMC,IAAKP,IAAOM;AAClB,eAASE,IAAI,GAAGA,IAAIL,GAAOK,KAAK;AAC5B,cAAMC,IAAKV,IAAOS;AAClB,YAAIE,IAAM;AACV,iBAASC,IAAK,GAAGA,IAAK3C,GAAc2C,KAAM;AACtC,gBAAMC,IAAOlD,EAAckB,IAAW+B,IAAKvC,IAAaW,CAAC,GACnD8B,IAAQjD,EAAc+C,IAAKhC,IAAa4B,IAAKrC,IAAQuC,CAAE;AAC7D,UAAAC,KAAOE,IAAOC;AAAA,QAClB;AACA,QAAAR,EAASC,IAAIH,IAAQK,CAAC,IAAIM,GAAQJ,CAAG;AAAA,MACzC;AAAA,IACJ;AAEA,UAAMK,IAAUC,GAAeX,GAAUF,GAAOC,GAAOZ,GAAOC,CAAK,GAC7DwB,IAAS,IAAI,WAAWzB,IAAQC,CAAK;AAC3C,aAASyB,IAAI,GAAGA,IAAID,EAAO,QAAQC;AAC/B,MAAAD,EAAOC,CAAC,IAAIH,EAAQG,CAAC,KAAK1C,IAAgB,MAAM;AAGpD,IAAAK,EAAQ,KAAK,EAAE,MAAAO,GAAM,SAAAE,GAAS,YAAAC,GAAY,MAAM,IAAIG,EAAKuB,GAAQzB,GAAOC,CAAK,EAAA,CAAG;AAAA,EACpF;AAEA,SAAOZ;AACX;AAEA,SAASiC,GAAQN,GAAmB;AAChC,MAAIA,KAAK;AACL,WAAO,KAAK,IAAI,KAAK,IAAI,CAACA,CAAC;AAE/B,QAAMW,IAAI,KAAK,IAAIX,CAAC;AACpB,SAAOW,KAAK,IAAIA;AACpB;AAEA,IAAIC,IAAyB;AAGtB,SAASC,GACZ3D,GACAC,GACAC,GACAC,GACAC,GACqB;AACrB,SAAKsD,MACDA,IAAyB,IACzB,QAAQ;AAAA,IACJ;AAAA,EAAA,IAID3D,GAAcC,GAAeC,GAAeC,GAAeC,GAAeC,CAAO;AAC5F;AAKA,SAASkD,GACLM,GACAC,GACAC,GACAC,GACAC,GACY;AACZ,QAAMC,IAAM,IAAI,aAAaF,IAAcC,CAAY;AACvD,MAAID,MAAgB,KAAKC,MAAiB,KAAKH,MAAa,KAAKC,MAAc;AAC3E,WAAOG;AAEX,MAAIF,MAAgBF,KAAYG,MAAiBF;AAC7C,WAAAG,EAAI,IAAIL,CAAG,GACJK;AAEX,QAAMC,IAAKL,IAAWE,GAChBI,IAAKL,IAAYE;AACvB,WAASpB,IAAI,GAAGA,IAAIoB,GAAcpB,KAAK;AACnC,UAAMwB,KAAMxB,IAAI,OAAOuB,IAAK,KACtBE,IAAK,KAAK,IAAI,GAAG,KAAK,MAAMD,CAAE,CAAC,GAC/B7C,IAAK,KAAK,IAAIuC,IAAY,GAAGO,IAAK,CAAC,GACnCC,IAAK,KAAK,IAAI,GAAG,KAAK,IAAI,GAAGF,IAAKC,CAAE,CAAC;AAC3C,aAAS,IAAI,GAAG,IAAIN,GAAa,KAAK;AAClC,YAAMQ,KAAM,IAAI,OAAOL,IAAK,KACtBM,IAAK,KAAK,IAAI,GAAG,KAAK,MAAMD,CAAE,CAAC,GAC/BjD,IAAK,KAAK,IAAIuC,IAAW,GAAGW,IAAK,CAAC,GAClCC,IAAK,KAAK,IAAI,GAAG,KAAK,IAAI,GAAGF,IAAKC,CAAE,CAAC,GACrCE,IAAMd,EAAIS,IAAKR,IAAWW,CAAE,GAC5BG,IAAMf,EAAIS,IAAKR,IAAWvC,CAAE,GAC5BsD,IAAMhB,EAAIrC,IAAKsC,IAAWW,CAAE,GAC5BK,IAAMjB,EAAIrC,IAAKsC,IAAWvC,CAAE,GAC5BwD,IAAMJ,KAAO,IAAID,KAAME,IAAMF,GAC7BM,IAAMH,KAAO,IAAIH,KAAMI,IAAMJ;AACnC,MAAAR,EAAIrB,IAAImB,IAAc,CAAC,IAAIe,KAAO,IAAIR,KAAMS,IAAMT;AAAA,IACtD;AAAA,EACJ;AACA,SAAOL;AACX;"}
+{"version":3,"file":"segmentation.js","names":[],"sources":["../../../src/vision/postprocess/segmentation.ts"],"sourcesContent":["/**\n * Segmentation head postprocessing: YOLO instance-segmentation decoding.\n *\n * Compatible with YOLOv8-seg / YOLOv11-seg exports (and any later seg head\n * sharing the layout). The model produces two output tensors:\n *\n * - `output0` of shape `(1, 4 + numClasses + numMaskCoefs, numAnchors)` — the\n *   same per-anchor predictions as plain YOLO detection plus an extra\n *   `numMaskCoefs` (typically 32) channels of mask coefficients.\n * - `output1` of shape `(1, numMaskCoefs, maskH, maskW)` — a set of \"prototype\"\n *   masks shared across all anchors.\n *\n * The per-anchor decode (xywh→xyxy, undo letterbox, per-class NMS, sort & cap)\n * is delegated to {@link decodeYoloAnchors}; this module only handles the\n * mask-specific work: matmul against prototypes, sigmoid, bilinear resize and\n * thresholding.\n */\n\nimport { decodeYoloAnchors } from \"./detection\";\nimport { BoundingBox, Mask } from \"../types\";\n\nexport interface DecodeYoloSegOptions {\n    readonly numClasses: number;\n    /** Model input `[width, height]` (post-letterbox). */\n    readonly inputWidth: number;\n    readonly inputHeight: number;\n    /** Original image `[width, height]`. */\n    readonly originalWidth: number;\n    readonly originalHeight: number;\n    /** Letterbox horizontal padding in input-tensor pixels. */\n    readonly padLeft: number;\n    /** Letterbox vertical padding in input-tensor pixels. */\n    readonly padTop: number;\n    /** Letterbox scale factor. */\n    readonly scale: number;\n    readonly confThreshold: number;\n    readonly iouThreshold: number;\n    readonly maxDetections: number;\n    /** Probability cutoff applied to the soft mask. Defaults to `0.5`. */\n    readonly maskThreshold?: number;\n}\n\nexport interface DecodedSegmentation {\n    readonly bbox: BoundingBox;\n    readonly classId: number;\n    readonly confidence: number;\n    /** Binary mask cropped to `bbox`. Width/height match `bbox.asIntXyxy()` extents. */\n    readonly mask: Mask;\n}\n\n/**\n * Decode YOLO segmentation raw outputs into a list of segmented instances.\n *\n * Compatible with YOLOv8-seg / YOLOv11-seg.\n *\n * @param perAnchorData Flat `output0`, length `(4 + numClasses + numMaskCoefs) * numAnchors`.\n * @param perAnchorDims Dims as reported by ORT, e.g. `[1, 116, 8400]`.\n * @param prototypeData Flat `output1`, length `numMaskCoefs * maskH * maskW`.\n * @param prototypeDims Dims as reported by ORT, e.g. `[1, 32, 160, 160]`.\n */\nexport function decodeYoloSeg(\n    perAnchorData: Float32Array,\n    perAnchorDims: readonly number[],\n    prototypeData: Float32Array,\n    prototypeDims: readonly number[],\n    options: DecodeYoloSegOptions,\n): DecodedSegmentation[] {\n    // Strip batch dim from prototypes; perAnchor batch is validated by the helper.\n    let pDims = prototypeDims;\n    if (pDims.length === 4) {\n        if (pDims[0] !== 1) {\n            throw new Error(`decodeYoloSeg: expected batch size 1 in prototypes, got ${pDims[0]}.`);\n        }\n        pDims = [pDims[1] as number, pDims[2] as number, pDims[3] as number];\n    }\n    if (pDims.length !== 3) {\n        throw new Error(\n            `decodeYoloSeg: expected 3-D prototypes after batch removal, got dims=${JSON.stringify(prototypeDims)}.`,\n        );\n    }\n    const numMaskCoefs = pDims[0] as number;\n    const maskH = pDims[1] as number;\n    const maskW = pDims[2] as number;\n\n    const channels = perAnchorDims.length === 3 ? perAnchorDims[1] : perAnchorDims[0];\n    const numAnchors = perAnchorDims.length === 3 ? perAnchorDims[2] : perAnchorDims[1];\n    if (channels === undefined || numAnchors === undefined) {\n        throw new Error(\n            `decodeYoloSeg: cannot read channels/numAnchors from dims=${JSON.stringify(perAnchorDims)}.`,\n        );\n    }\n\n    const expectedChannels = 4 + options.numClasses + numMaskCoefs;\n    if (channels !== expectedChannels) {\n        throw new Error(\n            `decodeYoloSeg: channels=${channels} does not match 4 + numClasses(${options.numClasses}) + numMaskCoefs(${numMaskCoefs}) = ${expectedChannels}.`,\n        );\n    }\n    if (prototypeData.length !== numMaskCoefs * maskH * maskW) {\n        throw new Error(\n            `decodeYoloSeg: prototype length ${prototypeData.length} does not match dims=${JSON.stringify(prototypeDims)}.`,\n        );\n    }\n\n    const decoded = decodeYoloAnchors(perAnchorData, perAnchorDims, {\n        numClasses: options.numClasses,\n        originalWidth: options.originalWidth,\n        originalHeight: options.originalHeight,\n        padLeft: options.padLeft,\n        padTop: options.padTop,\n        scale: options.scale,\n        confThreshold: options.confThreshold,\n        iouThreshold: options.iouThreshold,\n        maxDetections: options.maxDetections,\n    });\n\n    if (decoded.anchorIndices.length === 0) return [];\n\n    const maskThreshold = options.maskThreshold ?? 0.5;\n    const scaleX = maskW / options.inputWidth;\n    const scaleY = maskH / options.inputHeight;\n    const protoPlane = maskH * maskW;\n    const coefBase = (4 + options.numClasses) * numAnchors;\n\n    const results: DecodedSegmentation[] = [];\n\n    for (let i = 0; i < decoded.anchorIndices.length; i++) {\n        const a = decoded.anchorIndices[i] as number;\n        const x1 = decoded.boxesXyxy[i * 4] as number;\n        const y1 = decoded.boxesXyxy[i * 4 + 1] as number;\n        const x2 = decoded.boxesXyxy[i * 4 + 2] as number;\n        const y2 = decoded.boxesXyxy[i * 4 + 3] as number;\n        const bbox = new BoundingBox(x1, y1, x2, y2);\n        const classId = decoded.classIds[i] as number;\n        const confidence = decoded.confidences[i] as number;\n\n        const bboxW = Math.max(0, Math.trunc(x2) - Math.trunc(x1));\n        const bboxH = Math.max(0, Math.trunc(y2) - Math.trunc(y1));\n\n        if (bboxW === 0 || bboxH === 0) {\n            results.push({ bbox, classId, confidence, mask: new Mask(new Uint8Array(0), 0, 0) });\n            continue;\n        }\n\n        // Bbox in input-tensor coords, then in low-res mask coords.\n        const ibx1 = x1 * options.scale + options.padLeft;\n        const iby1 = y1 * options.scale + options.padTop;\n        const ibx2 = x2 * options.scale + options.padLeft;\n        const iby2 = y2 * options.scale + options.padTop;\n\n        const mbx1 = Math.max(0, Math.floor(ibx1 * scaleX));\n        const mby1 = Math.max(0, Math.floor(iby1 * scaleY));\n        const mbx2 = Math.min(maskW, Math.ceil(ibx2 * scaleX));\n        const mby2 = Math.min(maskH, Math.ceil(iby2 * scaleY));\n\n        if (mbx2 <= mbx1 || mby2 <= mby1) {\n            results.push({\n                bbox,\n                classId,\n                confidence,\n                mask: new Mask(new Uint8Array(bboxW * bboxH), bboxW, bboxH),\n            });\n            continue;\n        }\n\n        // Compute soft mask only within the prototype region under this bbox.\n        const cropW = mbx2 - mbx1;\n        const cropH = mby2 - mby1;\n        const softCrop = new Float32Array(cropW * cropH);\n        for (let y = 0; y < cropH; y++) {\n            const py = mby1 + y;\n            for (let x = 0; x < cropW; x++) {\n                const px = mbx1 + x;\n                let sum = 0;\n                for (let kk = 0; kk < numMaskCoefs; kk++) {\n                    const coef = perAnchorData[coefBase + kk * numAnchors + a];\n                    const proto = prototypeData[kk * protoPlane + py * maskW + px];\n                    sum += coef * proto;\n                }\n                softCrop[y * cropW + x] = sigmoid(sum);\n            }\n        }\n\n        const resized = resizeBilinear(softCrop, cropW, cropH, bboxW, bboxH);\n        const binary = new Uint8Array(bboxW * bboxH);\n        for (let j = 0; j < binary.length; j++) {\n            binary[j] = resized[j] >= maskThreshold ? 255 : 0;\n        }\n\n        results.push({ bbox, classId, confidence, mask: new Mask(binary, bboxW, bboxH) });\n    }\n\n    return results;\n}\n\nfunction sigmoid(x: number): number {\n    if (x >= 0) {\n        return 1 / (1 + Math.exp(-x));\n    }\n    const e = Math.exp(x);\n    return e / (1 + e);\n}\n\nlet _warnedDecodeYoloV8Seg = false;\n\n/** @deprecated since 0.2.0 — use {@link decodeYoloSeg}. Will be removed in 0.3.0. */\nexport function decodeYoloV8Seg(\n    perAnchorData: Float32Array,\n    perAnchorDims: readonly number[],\n    prototypeData: Float32Array,\n    prototypeDims: readonly number[],\n    options: DecodeYoloSegOptions,\n): DecodedSegmentation[] {\n    if (!_warnedDecodeYoloV8Seg) {\n        _warnedDecodeYoloV8Seg = true;\n        console.warn(\n            \"[@ort-vision-sdk/web] decodeYoloV8Seg is deprecated since 0.2.0; use decodeYoloSeg. \" +\n                \"The alias will be removed in 0.3.0.\",\n        );\n    }\n    return decodeYoloSeg(perAnchorData, perAnchorDims, prototypeData, prototypeDims, options);\n}\n\n/** @deprecated since 0.2.0 — use {@link DecodeYoloSegOptions}. */\nexport type DecodeYoloV8SegOptions = DecodeYoloSegOptions;\n\nfunction resizeBilinear(\n    src: Float32Array,\n    srcWidth: number,\n    srcHeight: number,\n    targetWidth: number,\n    targetHeight: number,\n): Float32Array {\n    const out = new Float32Array(targetWidth * targetHeight);\n    if (targetWidth === 0 || targetHeight === 0 || srcWidth === 0 || srcHeight === 0) {\n        return out;\n    }\n    if (targetWidth === srcWidth && targetHeight === srcHeight) {\n        out.set(src);\n        return out;\n    }\n    const sx = srcWidth / targetWidth;\n    const sy = srcHeight / targetHeight;\n    for (let y = 0; y < targetHeight; y++) {\n        const yy = (y + 0.5) * sy - 0.5;\n        const y0 = Math.max(0, Math.floor(yy));\n        const y1 = Math.min(srcHeight - 1, y0 + 1);\n        const wy = Math.max(0, Math.min(1, yy - y0));\n        for (let x = 0; x < targetWidth; x++) {\n            const xx = (x + 0.5) * sx - 0.5;\n            const x0 = Math.max(0, Math.floor(xx));\n            const x1 = Math.min(srcWidth - 1, x0 + 1);\n            const wx = Math.max(0, Math.min(1, xx - x0));\n            const v00 = src[y0 * srcWidth + x0];\n            const v01 = src[y0 * srcWidth + x1];\n            const v10 = src[y1 * srcWidth + x0];\n            const v11 = src[y1 * srcWidth + x1];\n            const top = v00 * (1 - wx) + v01 * wx;\n            const bot = v10 * (1 - wx) + v11 * wx;\n            out[y * targetWidth + x] = top * (1 - wy) + bot * wy;\n        }\n    }\n    return out;\n}\n"],"mappings":";;;AA4DA,SAAgB,EACZ,GACA,GACA,GACA,GACA,GACqB;CAErB,IAAI,IAAQ;CACZ,IAAI,EAAM,WAAW,GAAG;EACpB,IAAI,EAAM,OAAO,GACb,MAAU,MAAM,2DAA2D,EAAM,GAAG,EAAE;EAE1F,IAAQ;GAAC,EAAM;GAAc,EAAM;GAAc,EAAM;EAAY;CACvE;CACA,IAAI,EAAM,WAAW,GACjB,MAAU,MACN,wEAAwE,KAAK,UAAU,CAAa,EAAE,EAC1G;CAEJ,IAAM,IAAe,EAAM,IACrB,IAAQ,EAAM,IACd,IAAQ,EAAM,IAEd,IAAW,EAAc,WAAW,IAAI,EAAc,KAAK,EAAc,IACzE,IAAa,EAAc,WAAW,IAAI,EAAc,KAAK,EAAc;CACjF,IAAI,MAAa,KAAA,KAAa,MAAe,KAAA,GACzC,MAAU,MACN,4DAA4D,KAAK,UAAU,CAAa,EAAE,EAC9F;CAGJ,IAAM,IAAmB,IAAI,EAAQ,aAAa;CAClD,IAAI,MAAa,GACb,MAAU,MACN,2BAA2B,EAAS,iCAAiC,EAAQ,WAAW,mBAAmB,EAAa,MAAM,EAAiB,EACnJ;CAEJ,IAAI,EAAc,WAAW,IAAe,IAAQ,GAChD,MAAU,MACN,mCAAmC,EAAc,OAAO,uBAAuB,KAAK,UAAU,CAAa,EAAE,EACjH;CAGJ,IAAM,IAAU,EAAkB,GAAe,GAAe;EAC5D,YAAY,EAAQ;EACpB,eAAe,EAAQ;EACvB,gBAAgB,EAAQ;EACxB,SAAS,EAAQ;EACjB,QAAQ,EAAQ;EAChB,OAAO,EAAQ;EACf,eAAe,EAAQ;EACvB,cAAc,EAAQ;EACtB,eAAe,EAAQ;CAC3B,CAAC;CAED,IAAI,EAAQ,cAAc,WAAW,GAAG,OAAO,CAAC;CAEhD,IAAM,IAAgB,EAAQ,iBAAiB,IACzC,IAAS,IAAQ,EAAQ,YACzB,IAAS,IAAQ,EAAQ,aACzB,IAAa,IAAQ,GACrB,KAAY,IAAI,EAAQ,cAAc,GAEtC,IAAiC,CAAC;CAExC,KAAK,IAAI,IAAI,GAAG,IAAI,EAAQ,cAAc,QAAQ,KAAK;EACnD,IAAM,IAAI,EAAQ,cAAc,IAC1B,IAAK,EAAQ,UAAU,IAAI,IAC3B,IAAK,EAAQ,UAAU,IAAI,IAAI,IAC/B,IAAK,EAAQ,UAAU,IAAI,IAAI,IAC/B,IAAK,EAAQ,UAAU,IAAI,IAAI,IAC/B,IAAO,IAAI,EAAY,GAAI,GAAI,GAAI,CAAE,GACrC,IAAU,EAAQ,SAAS,IAC3B,IAAa,EAAQ,YAAY,IAEjC,IAAQ,KAAK,IAAI,GAAG,KAAK,MAAM,CAAE,IAAI,KAAK,MAAM,CAAE,CAAC,GACnD,IAAQ,KAAK,IAAI,GAAG,KAAK,MAAM,CAAE,IAAI,KAAK,MAAM,CAAE,CAAC;EAEzD,IAAI,MAAU,KAAK,MAAU,GAAG;GAC5B,EAAQ,KAAK;IAAE;IAAM;IAAS;IAAY,MAAM,IAAI,kBAAK,IAAI,WAAY,GAAG,GAAG,CAAC;GAAE,CAAC;GACnF;EACJ;EAGA,IAAM,IAAO,IAAK,EAAQ,QAAQ,EAAQ,SACpC,IAAO,IAAK,EAAQ,QAAQ,EAAQ,QACpC,IAAO,IAAK,EAAQ,QAAQ,EAAQ,SACpC,IAAO,IAAK,EAAQ,QAAQ,EAAQ,QAEpC,IAAO,KAAK,IAAI,GAAG,KAAK,MAAM,IAAO,CAAM,CAAC,GAC5C,IAAO,KAAK,IAAI,GAAG,KAAK,MAAM,IAAO,CAAM,CAAC,GAC5C,IAAO,KAAK,IAAI,GAAO,KAAK,KAAK,IAAO,CAAM,CAAC,GAC/C,IAAO,KAAK,IAAI,GAAO,KAAK,KAAK,IAAO,CAAM,CAAC;EAErD,IAAI,KAAQ,KAAQ,KAAQ,GAAM;GAC9B,EAAQ,KAAK;IACT;IACA;IACA;IACA,MAAM,IAAI,EAAK,IAAI,WAAW,IAAQ,CAAK,GAAG,GAAO,CAAK;GAC9D,CAAC;GACD;EACJ;EAGA,IAAM,IAAQ,IAAO,GACf,IAAQ,IAAO,GACf,IAAW,IAAI,aAAa,IAAQ,CAAK;EAC/C,KAAK,IAAI,IAAI,GAAG,IAAI,GAAO,KAAK;GAC5B,IAAM,IAAK,IAAO;GAClB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAO,KAAK;IAC5B,IAAM,IAAK,IAAO,GACd,IAAM;IACV,KAAK,IAAI,IAAK,GAAG,IAAK,GAAc,KAAM;KACtC,IAAM,IAAO,EAAc,IAAW,IAAK,IAAa,IAClD,IAAQ,EAAc,IAAK,IAAa,IAAK,IAAQ;KAC3D,KAAO,IAAO;IAClB;IACA,EAAS,IAAI,IAAQ,KAAK,EAAQ,CAAG;GACzC;EACJ;EAEA,IAAM,IAAU,EAAe,GAAU,GAAO,GAAO,GAAO,CAAK,GAC7D,IAAS,IAAI,WAAW,IAAQ,CAAK;EAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,EAAO,QAAQ,KAC/B,EAAO,KAAK,EAAQ,MAAM,IAAgB,MAAM;EAGpD,EAAQ,KAAK;GAAE;GAAM;GAAS;GAAY,MAAM,IAAI,EAAK,GAAQ,GAAO,CAAK;EAAE,CAAC;CACpF;CAEA,OAAO;AACX;AAEA,SAAS,EAAQ,GAAmB;CAChC,IAAI,KAAK,GACL,OAAO,KAAK,IAAI,KAAK,IAAI,CAAC,CAAC;CAE/B,IAAM,IAAI,KAAK,IAAI,CAAC;CACpB,OAAO,KAAK,IAAI;AACpB;AAEA,IAAI,IAAyB;AAG7B,SAAgB,EACZ,GACA,GACA,GACA,GACA,GACqB;CAQrB,OAPK,MACD,IAAyB,IACzB,QAAQ,KACJ,yHAEJ,IAEG,EAAc,GAAe,GAAe,GAAe,GAAe,CAAO;AAC5F;AAKA,SAAS,EACL,GACA,GACA,GACA,GACA,GACY;CACZ,IAAM,IAAM,IAAI,aAAa,IAAc,CAAY;CACvD,IAAI,MAAgB,KAAK,MAAiB,KAAK,MAAa,KAAK,MAAc,GAC3E,OAAO;CAEX,IAAI,MAAgB,KAAY,MAAiB,GAE7C,OADA,EAAI,IAAI,CAAG,GACJ;CAEX,IAAM,IAAK,IAAW,GAChB,IAAK,IAAY;CACvB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAc,KAAK;EACnC,IAAM,KAAM,IAAI,MAAO,IAAK,IACtB,IAAK,KAAK,IAAI,GAAG,KAAK,MAAM,CAAE,CAAC,GAC/B,IAAK,KAAK,IAAI,IAAY,GAAG,IAAK,CAAC,GACnC,IAAK,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,IAAK,CAAE,CAAC;EAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,GAAa,KAAK;GAClC,IAAM,KAAM,IAAI,MAAO,IAAK,IACtB,IAAK,KAAK,IAAI,GAAG,KAAK,MAAM,CAAE,CAAC,GAC/B,IAAK,KAAK,IAAI,IAAW,GAAG,IAAK,CAAC,GAClC,IAAK,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,IAAK,CAAE,CAAC,GACrC,IAAM,EAAI,IAAK,IAAW,IAC1B,IAAM,EAAI,IAAK,IAAW,IAC1B,IAAM,EAAI,IAAK,IAAW,IAC1B,IAAM,EAAI,IAAK,IAAW,IAC1B,IAAM,KAAO,IAAI,KAAM,IAAM,GAC7B,IAAM,KAAO,IAAI,KAAM,IAAM;GACnC,EAAI,IAAI,IAAc,KAAK,KAAO,IAAI,KAAM,IAAM;EACtD;CACJ;CACA,OAAO;AACX"}

package/dist/vision/preprocess/image.cjs

@@ -1,2 +1,2 @@
-"use strict";Object.defineProperty(exports,Symbol.toStringTag,{value:"Module"});const D=require("onnxruntime-web"),d=require("../core/canvas.cjs"),x=require("../types.cjs");function $(t){const o=Object.create(null,{[Symbol.toStringTag]:{value:"Module"}});if(t){for(const n in t)if(n!=="default"){const e=Object.getOwnPropertyDescriptor(t,n);Object.defineProperty(o,n,e.get?e:{enumerable:!0,get:()=>t[n]})}}return o.default=t,Object.freeze(o)}const I=$(D);function v(t,o,n){if(o<=0||n<=0)throw new Error(`Invalid resize target ${o}x${n}.`);if(o===t.width&&n===t.height)return t;const e=d.createCanvas(t.width,t.height);d.get2DContext(e).putImageData(d.rgbToImageData(t),0,0);const s=d.createCanvas(o,n),a=d.get2DContext(s);a.imageSmoothingEnabled=!0,a.imageSmoothingQuality="high",a.drawImage(e,0,0,o,n);const f=a.getImageData(0,0,o,n);return d.imageDataToRGB(f)}function O(t,o,n,e=1/255){const r=new Float32Array(t.data.length),s=t.data,a=o[0],f=o[1],c=o[2],w=n[0],y=n[1],i=n[2];for(let l=0;l<s.length;l+=3)r[l]=(s[l]*e-a)/w,r[l+1]=(s[l+1]*e-f)/y,r[l+2]=(s[l+2]*e-c)/i;return r}function b(t,o=1/255){const n=new Float32Array(t.data.length),e=t.data;for(let r=0;r<e.length;r++)n[r]=e[r]*o;return n}function T(t,o,n,e=3){const r=o*n*e;if(t.length!==r)throw new Error(`toCHW: expected length ${r} for ${o}x${n}x${e}, got ${t.length}.`);const s=new Float32Array(r),a=o*n;for(let f=0;f<n;f++)for(let c=0;c<o;c++){const w=(f*o+c)*e,y=f*o+c;for(let i=0;i<e;i++)s[i*a+y]=t[w+i]}return s}function z(t,o){return new I.Tensor("float32",t,o)}function F(t){const o=b(t);return T(o,t.width,t.height,3)}function M(t,o,n){if(t.length!==o*n*3)throw new Error(`fromCv2: data length ${t.length} does not match width * height * 3 = ${o*n*3}.`);const e=new Uint8Array(t.length);for(let r=0;r<t.length;r+=3)e[r]=t[r+2],e[r+1]=t[r+1],e[r+2]=t[r];return new x.RGBImage(e,o,n)}function A(t){const o=t.data,n=new Uint8Array(o.length);for(let e=0;e<o.length;e+=3)n[e]=o[e+2],n[e+1]=o[e+1],n[e+2]=o[e];return n}function S(t,o,n,e=[114,114,114]){const r=Math.min(o/t.width,n/t.height),s=Math.round(t.width*r),a=Math.round(t.height*r),f=v(t,s,a),c=new Uint8Array(o*n*3),w=e[0],y=e[1],i=e[2];for(let u=0;u<c.length;u+=3)c[u]=w,c[u+1]=y,c[u+2]=i;const l=Math.floor((o-s)/2),C=Math.floor((n-a)/2),p=s*3;for(let u=0;u<a;u++){const m=u*p,h=((C+u)*o+l)*3;c.set(f.data.subarray(m,m+p),h)}return{image:new x.RGBImage(c,o,n),scale:r,padLeft:l,padTop:C}}exports.fromCv2=M;exports.letterbox=S;exports.normalize=O;exports.resize=v;exports.toCHW=T;exports.toCv2=A;exports.toFloat32=b;exports.toFloat32Tensor=z;exports.toTensor=F;
-//# sourceMappingURL=image.cjs.map
+const e=require("../../_virtual/_rolldown/runtime.cjs"),t=require("../types.cjs"),n=require("../core/canvas.cjs");let r=require("onnxruntime-web");r=e.__toESM(r,1);function i(e,t,r){if(t<=0||r<=0)throw Error(`Invalid resize target ${t}x${r}.`);if(t===e.width&&r===e.height)return e;let i=n.createCanvas(e.width,e.height);n.get2DContext(i).putImageData(n.rgbToImageData(e),0,0);let a=n.get2DContext(n.createCanvas(t,r));return a.imageSmoothingEnabled=!0,a.imageSmoothingQuality=`high`,a.drawImage(i,0,0,t,r),n.imageDataToRGB(a.getImageData(0,0,t,r))}function a(e,t,n,r=1/255){let i=new Float32Array(e.data.length),a=e.data,o=t[0],s=t[1],c=t[2],l=n[0],u=n[1],d=n[2];for(let e=0;e<a.length;e+=3)i[e]=(a[e]*r-o)/l,i[e+1]=(a[e+1]*r-s)/u,i[e+2]=(a[e+2]*r-c)/d;return i}function o(e,t=1/255){let n=new Float32Array(e.data.length),r=e.data;for(let e=0;e<r.length;e++)n[e]=r[e]*t;return n}function s(e,t,n,r=3){let i=t*n*r;if(e.length!==i)throw Error(`toCHW: expected length ${i} for ${t}x${n}x${r}, got ${e.length}.`);let a=new Float32Array(i),o=t*n;for(let i=0;i<n;i++)for(let n=0;n<t;n++){let s=(i*t+n)*r,c=i*t+n;for(let t=0;t<r;t++)a[t*o+c]=e[s+t]}return a}function c(e,t){return new r.Tensor(`float32`,e,t)}function l(e){return s(o(e),e.width,e.height,3)}function u(e,n,r){if(e.length!==n*r*3)throw Error(`fromCv2: data length ${e.length} does not match width * height * 3 = ${n*r*3}.`);let i=new Uint8Array(e.length);for(let t=0;t<e.length;t+=3)i[t]=e[t+2],i[t+1]=e[t+1],i[t+2]=e[t];return new t.RGBImage(i,n,r)}function d(e){let t=e.data,n=new Uint8Array(t.length);for(let e=0;e<t.length;e+=3)n[e]=t[e+2],n[e+1]=t[e+1],n[e+2]=t[e];return n}function f(e,n,r,a=[114,114,114]){let o=Math.min(n/e.width,r/e.height),s=Math.round(e.width*o),c=Math.round(e.height*o),l=i(e,s,c),u=new Uint8Array(n*r*3),d=a[0],f=a[1],p=a[2];for(let e=0;e<u.length;e+=3)u[e]=d,u[e+1]=f,u[e+2]=p;let m=Math.floor((n-s)/2),h=Math.floor((r-c)/2),g=s*3;for(let e=0;e<c;e++){let t=e*g,r=((h+e)*n+m)*3;u.set(l.data.subarray(t,t+g),r)}return{image:new t.RGBImage(u,n,r),scale:o,padLeft:m,padTop:h}}exports.fromCv2=u,exports.letterbox=f,exports.normalize=a,exports.resize=i,exports.toCHW=s,exports.toCv2=d,exports.toFloat32=o,exports.toFloat32Tensor=c,exports.toTensor=l;
+//# sourceMappingURL=image.cjs.map

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

@@ -1 +1 @@
-{"version":3,"file":"image.cjs","sources":["../../../src/vision/preprocess/image.ts"],"sourcesContent":["/**\n * Composable image preprocessing primitives (resize, normalize, letterbox, layout).\n *\n * Mirrors the Python `preprocess.image` module. Operates on the canonical\n * {@link RGBImage} (HWC RGB uint8) and produces uint8 or float32 buffers\n * depending on the operation.\n */\n\nimport * as ortRuntime from \"onnxruntime-web\";\nimport type * as ort from \"onnxruntime-web\";\n\nimport { createCanvas, get2DContext, imageDataToRGB, rgbToImageData } from \"../core/canvas\";\nimport { RGBImage } from \"../types\";\n\n/** Resize an image to `(targetWidth, targetHeight)` using high-quality canvas resampling. */\nexport function resize(image: RGBImage, targetWidth: number, targetHeight: number): RGBImage {\n    if (targetWidth <= 0 || targetHeight <= 0) {\n        throw new Error(`Invalid resize target ${targetWidth}x${targetHeight}.`);\n    }\n    if (targetWidth === image.width && targetHeight === image.height) {\n        return image;\n    }\n\n    const srcCanvas = createCanvas(image.width, image.height);\n    const srcCtx = get2DContext(srcCanvas);\n    srcCtx.putImageData(rgbToImageData(image), 0, 0);\n\n    const dstCanvas = createCanvas(targetWidth, targetHeight);\n    const dstCtx = get2DContext(dstCanvas);\n    dstCtx.imageSmoothingEnabled = true;\n    dstCtx.imageSmoothingQuality = \"high\";\n    dstCtx.drawImage(srcCanvas as CanvasImageSource, 0, 0, targetWidth, targetHeight);\n\n    const imageData = dstCtx.getImageData(0, 0, targetWidth, targetHeight);\n    return imageDataToRGB(imageData);\n}\n\n/**\n * Convert a uint8 image to a normalized float32 array (HWC layout preserved).\n *\n * Applies `(pixel * scale - mean) / std` channel-wise.\n */\nexport function normalize(\n    image: RGBImage,\n    mean: readonly [number, number, number],\n    std: readonly [number, number, number],\n    scale: number = 1 / 255,\n): Float32Array {\n    const out = new Float32Array(image.data.length);\n    const data = image.data;\n    const m0 = mean[0];\n    const m1 = mean[1];\n    const m2 = mean[2];\n    const s0 = std[0];\n    const s1 = std[1];\n    const s2 = std[2];\n    for (let i = 0; i < data.length; i += 3) {\n        out[i] = ((data[i] as number) * scale - m0) / s0;\n        out[i + 1] = ((data[i + 1] as number) * scale - m1) / s1;\n        out[i + 2] = ((data[i + 2] as number) * scale - m2) / s2;\n    }\n    return out;\n}\n\n/** Convert a uint8 image to a `Float32Array` in `[0, 1]` (HWC layout preserved). */\nexport function toFloat32(image: RGBImage, scale: number = 1 / 255): Float32Array {\n    const out = new Float32Array(image.data.length);\n    const data = image.data;\n    for (let i = 0; i < data.length; i++) {\n        out[i] = (data[i] as number) * scale;\n    }\n    return out;\n}\n\n/**\n * Transpose interleaved HWC data to planar CHW layout.\n *\n * @param hwc Source array of length `width * height * channels`.\n */\nexport function toCHW(\n    hwc: Float32Array,\n    width: number,\n    height: number,\n    channels: number = 3,\n): Float32Array {\n    const expected = width * height * channels;\n    if (hwc.length !== expected) {\n        throw new Error(\n            `toCHW: expected length ${expected} for ${width}x${height}x${channels}, got ${hwc.length}.`,\n        );\n    }\n    const chw = new Float32Array(expected);\n    const plane = width * height;\n    for (let y = 0; y < height; y++) {\n        for (let x = 0; x < width; x++) {\n            const hwcBase = (y * width + x) * channels;\n            const planeIdx = y * width + x;\n            for (let c = 0; c < channels; c++) {\n                chw[c * plane + planeIdx] = hwc[hwcBase + c] as number;\n            }\n        }\n    }\n    return chw;\n}\n\n/** Wrap a Float32 buffer into an `ort.Tensor`. */\nexport function toFloat32Tensor(data: Float32Array, dims: readonly number[]): ort.Tensor {\n    return new ortRuntime.Tensor(\"float32\", data, dims as number[]);\n}\n\n/**\n * Convert an HWC uint8 image to a CHW `Float32Array` scaled to `[0, 1]`.\n *\n * Mirrors `torchvision.transforms.ToTensor()` semantics: HWC → CHW,\n * `uint8 → float32 / 255`. Useful as input to YOLO-style detectors that\n * don't require ImageNet normalization.\n *\n * @returns CHW `Float32Array` of length `width * height * 3`.\n */\nexport function toTensor(image: RGBImage): Float32Array {\n    const f32 = toFloat32(image);\n    return toCHW(f32, image.width, image.height, 3);\n}\n\n/**\n * Convert an HWC BGR uint8 buffer (OpenCV layout) to the SDK's HWC RGB.\n *\n * Use when you receive image bytes from `cv2.imencode` over the wire and\n * want to feed them to the SDK without going through a canvas decode.\n *\n * @param bgr Flat BGR Uint8Array of length `width * height * 3`.\n */\nexport function fromCv2(bgr: Uint8Array, width: number, height: number): RGBImage {\n    if (bgr.length !== width * height * 3) {\n        throw new Error(\n            `fromCv2: data length ${bgr.length} does not match width * height * 3 = ${width * height * 3}.`,\n        );\n    }\n    const rgb = new Uint8Array(bgr.length);\n    for (let i = 0; i < bgr.length; i += 3) {\n        rgb[i] = bgr[i + 2] as number;\n        rgb[i + 1] = bgr[i + 1] as number;\n        rgb[i + 2] = bgr[i] as number;\n    }\n    return new RGBImage(rgb, width, height);\n}\n\n/**\n * Convert the SDK's HWC RGB image to an HWC BGR `Uint8Array` (OpenCV layout).\n *\n * Useful for round-tripping data to a Python OpenCV consumer.\n */\nexport function toCv2(image: RGBImage): Uint8Array {\n    const rgb = image.data;\n    const bgr = new Uint8Array(rgb.length);\n    for (let i = 0; i < rgb.length; i += 3) {\n        bgr[i] = rgb[i + 2] as number;\n        bgr[i + 1] = rgb[i + 1] as number;\n        bgr[i + 2] = rgb[i] as number;\n    }\n    return bgr;\n}\n\nexport interface LetterboxResult {\n    /** The padded image at the target size. */\n    readonly image: RGBImage;\n    /** The factor applied to the original image (`< 1` if downscaled). */\n    readonly scale: number;\n    /** Horizontal padding in pixels (left side; right side has the same or +1). */\n    readonly padLeft: number;\n    /** Vertical padding in pixels (top side). */\n    readonly padTop: number;\n}\n\n/**\n * Resize preserving aspect ratio, padding to `(targetWidth, targetHeight)`\n * with a constant fill color.\n *\n * Standard YOLO preprocessing — returning `scale` and `padLeft`/`padTop`\n * lets callers map detections back to the original image coordinates.\n */\nexport function letterbox(\n    image: RGBImage,\n    targetWidth: number,\n    targetHeight: number,\n    fill: readonly [number, number, number] = [114, 114, 114],\n): LetterboxResult {\n    const scale = Math.min(targetWidth / image.width, targetHeight / image.height);\n    const newW = Math.round(image.width * scale);\n    const newH = Math.round(image.height * scale);\n    const resized = resize(image, newW, newH);\n\n    const out = new Uint8Array(targetWidth * targetHeight * 3);\n    const f0 = fill[0];\n    const f1 = fill[1];\n    const f2 = fill[2];\n    for (let i = 0; i < out.length; i += 3) {\n        out[i] = f0;\n        out[i + 1] = f1;\n        out[i + 2] = f2;\n    }\n\n    const padLeft = Math.floor((targetWidth - newW) / 2);\n    const padTop = Math.floor((targetHeight - newH) / 2);\n    const rowBytes = newW * 3;\n    for (let y = 0; y < newH; y++) {\n        const srcOffset = y * rowBytes;\n        const dstOffset = ((padTop + y) * targetWidth + padLeft) * 3;\n        out.set(resized.data.subarray(srcOffset, srcOffset + rowBytes), dstOffset);\n    }\n\n    return {\n        image: new RGBImage(out, targetWidth, targetHeight),\n        scale,\n        padLeft,\n        padTop,\n    };\n}\n"],"names":["resize","image","targetWidth","targetHeight","srcCanvas","createCanvas","get2DContext","rgbToImageData","dstCanvas","dstCtx","imageData","imageDataToRGB","normalize","mean","std","scale","out","data","m0","m1","m2","s0","s1","s2","i","toFloat32","toCHW","hwc","width","height","channels","expected","chw","plane","y","x","hwcBase","planeIdx","c","toFloat32Tensor","dims","ortRuntime","toTensor","f32","fromCv2","bgr","rgb","RGBImage","toCv2","letterbox","fill","newW","newH","resized","f0","f1","f2","padLeft","padTop","rowBytes","srcOffset","dstOffset"],"mappings":"ucAeO,SAASA,EAAOC,EAAiBC,EAAqBC,EAAgC,CACzF,GAAID,GAAe,GAAKC,GAAgB,EACpC,MAAM,IAAI,MAAM,yBAAyBD,CAAW,IAAIC,CAAY,GAAG,EAE3E,GAAID,IAAgBD,EAAM,OAASE,IAAiBF,EAAM,OACtD,OAAOA,EAGX,MAAMG,EAAYC,EAAAA,aAAaJ,EAAM,MAAOA,EAAM,MAAM,EACzCK,EAAAA,aAAaF,CAAS,EAC9B,aAAaG,EAAAA,eAAeN,CAAK,EAAG,EAAG,CAAC,EAE/C,MAAMO,EAAYH,EAAAA,aAAaH,EAAaC,CAAY,EAClDM,EAASH,EAAAA,aAAaE,CAAS,EACrCC,EAAO,sBAAwB,GAC/BA,EAAO,sBAAwB,OAC/BA,EAAO,UAAUL,EAAgC,EAAG,EAAGF,EAAaC,CAAY,EAEhF,MAAMO,EAAYD,EAAO,aAAa,EAAG,EAAGP,EAAaC,CAAY,EACrE,OAAOQ,EAAAA,eAAeD,CAAS,CACnC,CAOO,SAASE,EACZX,EACAY,EACAC,EACAC,EAAgB,EAAI,IACR,CACZ,MAAMC,EAAM,IAAI,aAAaf,EAAM,KAAK,MAAM,EACxCgB,EAAOhB,EAAM,KACbiB,EAAKL,EAAK,CAAC,EACXM,EAAKN,EAAK,CAAC,EACXO,EAAKP,EAAK,CAAC,EACXQ,EAAKP,EAAI,CAAC,EACVQ,EAAKR,EAAI,CAAC,EACVS,EAAKT,EAAI,CAAC,EAChB,QAASU,EAAI,EAAGA,EAAIP,EAAK,OAAQO,GAAK,EAClCR,EAAIQ,CAAC,GAAMP,EAAKO,CAAC,EAAeT,EAAQG,GAAMG,EAC9CL,EAAIQ,EAAI,CAAC,GAAMP,EAAKO,EAAI,CAAC,EAAeT,EAAQI,GAAMG,EACtDN,EAAIQ,EAAI,CAAC,GAAMP,EAAKO,EAAI,CAAC,EAAeT,EAAQK,GAAMG,EAE1D,OAAOP,CACX,CAGO,SAASS,EAAUxB,EAAiBc,EAAgB,EAAI,IAAmB,CAC9E,MAAMC,EAAM,IAAI,aAAaf,EAAM,KAAK,MAAM,EACxCgB,EAAOhB,EAAM,KACnB,QAASuB,EAAI,EAAGA,EAAIP,EAAK,OAAQO,IAC7BR,EAAIQ,CAAC,EAAKP,EAAKO,CAAC,EAAeT,EAEnC,OAAOC,CACX,CAOO,SAASU,EACZC,EACAC,EACAC,EACAC,EAAmB,EACP,CACZ,MAAMC,EAAWH,EAAQC,EAASC,EAClC,GAAIH,EAAI,SAAWI,EACf,MAAM,IAAI,MACN,0BAA0BA,CAAQ,QAAQH,CAAK,IAAIC,CAAM,IAAIC,CAAQ,SAASH,EAAI,MAAM,GAAA,EAGhG,MAAMK,EAAM,IAAI,aAAaD,CAAQ,EAC/BE,EAAQL,EAAQC,EACtB,QAASK,EAAI,EAAGA,EAAIL,EAAQK,IACxB,QAASC,EAAI,EAAGA,EAAIP,EAAOO,IAAK,CAC5B,MAAMC,GAAWF,EAAIN,EAAQO,GAAKL,EAC5BO,EAAWH,EAAIN,EAAQO,EAC7B,QAASG,EAAI,EAAGA,EAAIR,EAAUQ,IAC1BN,EAAIM,EAAIL,EAAQI,CAAQ,EAAIV,EAAIS,EAAUE,CAAC,CAEnD,CAEJ,OAAON,CACX,CAGO,SAASO,EAAgBtB,EAAoBuB,EAAqC,CACrF,OAAO,IAAIC,EAAW,OAAO,UAAWxB,EAAMuB,CAAgB,CAClE,CAWO,SAASE,EAASzC,EAA+B,CACpD,MAAM0C,EAAMlB,EAAUxB,CAAK,EAC3B,OAAOyB,EAAMiB,EAAK1C,EAAM,MAAOA,EAAM,OAAQ,CAAC,CAClD,CAUO,SAAS2C,EAAQC,EAAiBjB,EAAeC,EAA0B,CAC9E,GAAIgB,EAAI,SAAWjB,EAAQC,EAAS,EAChC,MAAM,IAAI,MACN,wBAAwBgB,EAAI,MAAM,wCAAwCjB,EAAQC,EAAS,CAAC,GAAA,EAGpG,MAAMiB,EAAM,IAAI,WAAWD,EAAI,MAAM,EACrC,QAASrB,EAAI,EAAGA,EAAIqB,EAAI,OAAQrB,GAAK,EACjCsB,EAAItB,CAAC,EAAIqB,EAAIrB,EAAI,CAAC,EAClBsB,EAAItB,EAAI,CAAC,EAAIqB,EAAIrB,EAAI,CAAC,EACtBsB,EAAItB,EAAI,CAAC,EAAIqB,EAAIrB,CAAC,EAEtB,OAAO,IAAIuB,EAAAA,SAASD,EAAKlB,EAAOC,CAAM,CAC1C,CAOO,SAASmB,EAAM/C,EAA6B,CAC/C,MAAM6C,EAAM7C,EAAM,KACZ4C,EAAM,IAAI,WAAWC,EAAI,MAAM,EACrC,QAAStB,EAAI,EAAGA,EAAIsB,EAAI,OAAQtB,GAAK,EACjCqB,EAAIrB,CAAC,EAAIsB,EAAItB,EAAI,CAAC,EAClBqB,EAAIrB,EAAI,CAAC,EAAIsB,EAAItB,EAAI,CAAC,EACtBqB,EAAIrB,EAAI,CAAC,EAAIsB,EAAItB,CAAC,EAEtB,OAAOqB,CACX,CAoBO,SAASI,EACZhD,EACAC,EACAC,EACA+C,EAA0C,CAAC,IAAK,IAAK,GAAG,EACzC,CACf,MAAMnC,EAAQ,KAAK,IAAIb,EAAcD,EAAM,MAAOE,EAAeF,EAAM,MAAM,EACvEkD,EAAO,KAAK,MAAMlD,EAAM,MAAQc,CAAK,EACrCqC,EAAO,KAAK,MAAMnD,EAAM,OAASc,CAAK,EACtCsC,EAAUrD,EAAOC,EAAOkD,EAAMC,CAAI,EAElCpC,EAAM,IAAI,WAAWd,EAAcC,EAAe,CAAC,EACnDmD,EAAKJ,EAAK,CAAC,EACXK,EAAKL,EAAK,CAAC,EACXM,EAAKN,EAAK,CAAC,EACjB,QAAS1B,EAAI,EAAGA,EAAIR,EAAI,OAAQQ,GAAK,EACjCR,EAAIQ,CAAC,EAAI8B,EACTtC,EAAIQ,EAAI,CAAC,EAAI+B,EACbvC,EAAIQ,EAAI,CAAC,EAAIgC,EAGjB,MAAMC,EAAU,KAAK,OAAOvD,EAAciD,GAAQ,CAAC,EAC7CO,EAAS,KAAK,OAAOvD,EAAeiD,GAAQ,CAAC,EAC7CO,EAAWR,EAAO,EACxB,QAASjB,EAAI,EAAGA,EAAIkB,EAAMlB,IAAK,CAC3B,MAAM0B,EAAY1B,EAAIyB,EAChBE,IAAcH,EAASxB,GAAKhC,EAAcuD,GAAW,EAC3DzC,EAAI,IAAIqC,EAAQ,KAAK,SAASO,EAAWA,EAAYD,CAAQ,EAAGE,CAAS,CAC7E,CAEA,MAAO,CACH,MAAO,IAAId,EAAAA,SAAS/B,EAAKd,EAAaC,CAAY,EAClD,MAAAY,EACA,QAAA0C,EACA,OAAAC,CAAA,CAER"}
+{"version":3,"file":"image.cjs","names":[],"sources":["../../../src/vision/preprocess/image.ts"],"sourcesContent":["/**\n * Composable image preprocessing primitives (resize, normalize, letterbox, layout).\n *\n * Mirrors the Python `preprocess.image` module. Operates on the canonical\n * {@link RGBImage} (HWC RGB uint8) and produces uint8 or float32 buffers\n * depending on the operation.\n */\n\nimport * as ortRuntime from \"onnxruntime-web\";\nimport type * as ort from \"onnxruntime-web\";\n\nimport { createCanvas, get2DContext, imageDataToRGB, rgbToImageData } from \"../core/canvas\";\nimport { RGBImage } from \"../types\";\n\n/** Resize an image to `(targetWidth, targetHeight)` using high-quality canvas resampling. */\nexport function resize(image: RGBImage, targetWidth: number, targetHeight: number): RGBImage {\n    if (targetWidth <= 0 || targetHeight <= 0) {\n        throw new Error(`Invalid resize target ${targetWidth}x${targetHeight}.`);\n    }\n    if (targetWidth === image.width && targetHeight === image.height) {\n        return image;\n    }\n\n    const srcCanvas = createCanvas(image.width, image.height);\n    const srcCtx = get2DContext(srcCanvas);\n    srcCtx.putImageData(rgbToImageData(image), 0, 0);\n\n    const dstCanvas = createCanvas(targetWidth, targetHeight);\n    const dstCtx = get2DContext(dstCanvas);\n    dstCtx.imageSmoothingEnabled = true;\n    dstCtx.imageSmoothingQuality = \"high\";\n    dstCtx.drawImage(srcCanvas as CanvasImageSource, 0, 0, targetWidth, targetHeight);\n\n    const imageData = dstCtx.getImageData(0, 0, targetWidth, targetHeight);\n    return imageDataToRGB(imageData);\n}\n\n/**\n * Convert a uint8 image to a normalized float32 array (HWC layout preserved).\n *\n * Applies `(pixel * scale - mean) / std` channel-wise.\n */\nexport function normalize(\n    image: RGBImage,\n    mean: readonly [number, number, number],\n    std: readonly [number, number, number],\n    scale: number = 1 / 255,\n): Float32Array {\n    const out = new Float32Array(image.data.length);\n    const data = image.data;\n    const m0 = mean[0];\n    const m1 = mean[1];\n    const m2 = mean[2];\n    const s0 = std[0];\n    const s1 = std[1];\n    const s2 = std[2];\n    for (let i = 0; i < data.length; i += 3) {\n        out[i] = ((data[i] as number) * scale - m0) / s0;\n        out[i + 1] = ((data[i + 1] as number) * scale - m1) / s1;\n        out[i + 2] = ((data[i + 2] as number) * scale - m2) / s2;\n    }\n    return out;\n}\n\n/** Convert a uint8 image to a `Float32Array` in `[0, 1]` (HWC layout preserved). */\nexport function toFloat32(image: RGBImage, scale: number = 1 / 255): Float32Array {\n    const out = new Float32Array(image.data.length);\n    const data = image.data;\n    for (let i = 0; i < data.length; i++) {\n        out[i] = (data[i] as number) * scale;\n    }\n    return out;\n}\n\n/**\n * Transpose interleaved HWC data to planar CHW layout.\n *\n * @param hwc Source array of length `width * height * channels`.\n */\nexport function toCHW(\n    hwc: Float32Array,\n    width: number,\n    height: number,\n    channels: number = 3,\n): Float32Array {\n    const expected = width * height * channels;\n    if (hwc.length !== expected) {\n        throw new Error(\n            `toCHW: expected length ${expected} for ${width}x${height}x${channels}, got ${hwc.length}.`,\n        );\n    }\n    const chw = new Float32Array(expected);\n    const plane = width * height;\n    for (let y = 0; y < height; y++) {\n        for (let x = 0; x < width; x++) {\n            const hwcBase = (y * width + x) * channels;\n            const planeIdx = y * width + x;\n            for (let c = 0; c < channels; c++) {\n                chw[c * plane + planeIdx] = hwc[hwcBase + c] as number;\n            }\n        }\n    }\n    return chw;\n}\n\n/** Wrap a Float32 buffer into an `ort.Tensor`. */\nexport function toFloat32Tensor(data: Float32Array, dims: readonly number[]): ort.Tensor {\n    return new ortRuntime.Tensor(\"float32\", data, dims as number[]);\n}\n\n/**\n * Convert an HWC uint8 image to a CHW `Float32Array` scaled to `[0, 1]`.\n *\n * Mirrors `torchvision.transforms.ToTensor()` semantics: HWC → CHW,\n * `uint8 → float32 / 255`. Useful as input to YOLO-style detectors that\n * don't require ImageNet normalization.\n *\n * @returns CHW `Float32Array` of length `width * height * 3`.\n */\nexport function toTensor(image: RGBImage): Float32Array {\n    const f32 = toFloat32(image);\n    return toCHW(f32, image.width, image.height, 3);\n}\n\n/**\n * Convert an HWC BGR uint8 buffer (OpenCV layout) to the SDK's HWC RGB.\n *\n * Use when you receive image bytes from `cv2.imencode` over the wire and\n * want to feed them to the SDK without going through a canvas decode.\n *\n * @param bgr Flat BGR Uint8Array of length `width * height * 3`.\n */\nexport function fromCv2(bgr: Uint8Array, width: number, height: number): RGBImage {\n    if (bgr.length !== width * height * 3) {\n        throw new Error(\n            `fromCv2: data length ${bgr.length} does not match width * height * 3 = ${width * height * 3}.`,\n        );\n    }\n    const rgb = new Uint8Array(bgr.length);\n    for (let i = 0; i < bgr.length; i += 3) {\n        rgb[i] = bgr[i + 2] as number;\n        rgb[i + 1] = bgr[i + 1] as number;\n        rgb[i + 2] = bgr[i] as number;\n    }\n    return new RGBImage(rgb, width, height);\n}\n\n/**\n * Convert the SDK's HWC RGB image to an HWC BGR `Uint8Array` (OpenCV layout).\n *\n * Useful for round-tripping data to a Python OpenCV consumer.\n */\nexport function toCv2(image: RGBImage): Uint8Array {\n    const rgb = image.data;\n    const bgr = new Uint8Array(rgb.length);\n    for (let i = 0; i < rgb.length; i += 3) {\n        bgr[i] = rgb[i + 2] as number;\n        bgr[i + 1] = rgb[i + 1] as number;\n        bgr[i + 2] = rgb[i] as number;\n    }\n    return bgr;\n}\n\nexport interface LetterboxResult {\n    /** The padded image at the target size. */\n    readonly image: RGBImage;\n    /** The factor applied to the original image (`< 1` if downscaled). */\n    readonly scale: number;\n    /** Horizontal padding in pixels (left side; right side has the same or +1). */\n    readonly padLeft: number;\n    /** Vertical padding in pixels (top side). */\n    readonly padTop: number;\n}\n\n/**\n * Resize preserving aspect ratio, padding to `(targetWidth, targetHeight)`\n * with a constant fill color.\n *\n * Standard YOLO preprocessing — returning `scale` and `padLeft`/`padTop`\n * lets callers map detections back to the original image coordinates.\n */\nexport function letterbox(\n    image: RGBImage,\n    targetWidth: number,\n    targetHeight: number,\n    fill: readonly [number, number, number] = [114, 114, 114],\n): LetterboxResult {\n    const scale = Math.min(targetWidth / image.width, targetHeight / image.height);\n    const newW = Math.round(image.width * scale);\n    const newH = Math.round(image.height * scale);\n    const resized = resize(image, newW, newH);\n\n    const out = new Uint8Array(targetWidth * targetHeight * 3);\n    const f0 = fill[0];\n    const f1 = fill[1];\n    const f2 = fill[2];\n    for (let i = 0; i < out.length; i += 3) {\n        out[i] = f0;\n        out[i + 1] = f1;\n        out[i + 2] = f2;\n    }\n\n    const padLeft = Math.floor((targetWidth - newW) / 2);\n    const padTop = Math.floor((targetHeight - newH) / 2);\n    const rowBytes = newW * 3;\n    for (let y = 0; y < newH; y++) {\n        const srcOffset = y * rowBytes;\n        const dstOffset = ((padTop + y) * targetWidth + padLeft) * 3;\n        out.set(resized.data.subarray(srcOffset, srcOffset + rowBytes), dstOffset);\n    }\n\n    return {\n        image: new RGBImage(out, targetWidth, targetHeight),\n        scale,\n        padLeft,\n        padTop,\n    };\n}\n"],"mappings":"oKAeA,SAAgB,EAAO,EAAiB,EAAqB,EAAgC,CACzF,GAAI,GAAe,GAAK,GAAgB,EACpC,MAAU,MAAM,yBAAyB,EAAY,GAAG,EAAa,EAAE,EAE3E,GAAI,IAAgB,EAAM,OAAS,IAAiB,EAAM,OACtD,OAAO,EAGX,IAAM,EAAY,EAAA,aAAa,EAAM,MAAO,EAAM,MAAM,EAExD,EADe,aAAa,CAC5B,CAAA,CAAO,aAAa,EAAA,eAAe,CAAK,EAAG,EAAG,CAAC,EAG/C,IAAM,EAAS,EAAA,aADG,EAAA,aAAa,EAAa,CAChB,CAAS,EAMrC,MALA,GAAO,sBAAwB,GAC/B,EAAO,sBAAwB,OAC/B,EAAO,UAAU,EAAgC,EAAG,EAAG,EAAa,CAAY,EAGzE,EAAA,eADW,EAAO,aAAa,EAAG,EAAG,EAAa,CACnC,CAAS,CACnC,CAOA,SAAgB,EACZ,EACA,EACA,EACA,EAAgB,EAAI,IACR,CACZ,IAAM,EAAM,IAAI,aAAa,EAAM,KAAK,MAAM,EACxC,EAAO,EAAM,KACb,EAAK,EAAK,GACV,EAAK,EAAK,GACV,EAAK,EAAK,GACV,EAAK,EAAI,GACT,EAAK,EAAI,GACT,EAAK,EAAI,GACf,IAAK,IAAI,EAAI,EAAG,EAAI,EAAK,OAAQ,GAAK,EAClC,EAAI,IAAO,EAAK,GAAgB,EAAQ,GAAM,EAC9C,EAAI,EAAI,IAAO,EAAK,EAAI,GAAgB,EAAQ,GAAM,EACtD,EAAI,EAAI,IAAO,EAAK,EAAI,GAAgB,EAAQ,GAAM,EAE1D,OAAO,CACX,CAGA,SAAgB,EAAU,EAAiB,EAAgB,EAAI,IAAmB,CAC9E,IAAM,EAAM,IAAI,aAAa,EAAM,KAAK,MAAM,EACxC,EAAO,EAAM,KACnB,IAAK,IAAI,EAAI,EAAG,EAAI,EAAK,OAAQ,IAC7B,EAAI,GAAM,EAAK,GAAgB,EAEnC,OAAO,CACX,CAOA,SAAgB,EACZ,EACA,EACA,EACA,EAAmB,EACP,CACZ,IAAM,EAAW,EAAQ,EAAS,EAClC,GAAI,EAAI,SAAW,EACf,MAAU,MACN,0BAA0B,EAAS,OAAO,EAAM,GAAG,EAAO,GAAG,EAAS,QAAQ,EAAI,OAAO,EAC7F,EAEJ,IAAM,EAAM,IAAI,aAAa,CAAQ,EAC/B,EAAQ,EAAQ,EACtB,IAAK,IAAI,EAAI,EAAG,EAAI,EAAQ,IACxB,IAAK,IAAI,EAAI,EAAG,EAAI,EAAO,IAAK,CAC5B,IAAM,GAAW,EAAI,EAAQ,GAAK,EAC5B,EAAW,EAAI,EAAQ,EAC7B,IAAK,IAAI,EAAI,EAAG,EAAI,EAAU,IAC1B,EAAI,EAAI,EAAQ,GAAY,EAAI,EAAU,EAElD,CAEJ,OAAO,CACX,CAGA,SAAgB,EAAgB,EAAoB,EAAqC,CACrF,OAAO,IAAI,EAAW,OAAO,UAAW,EAAM,CAAgB,CAClE,CAWA,SAAgB,EAAS,EAA+B,CAEpD,OAAO,EADK,EAAU,CACT,EAAK,EAAM,MAAO,EAAM,OAAQ,CAAC,CAClD,CAUA,SAAgB,EAAQ,EAAiB,EAAe,EAA0B,CAC9E,GAAI,EAAI,SAAW,EAAQ,EAAS,EAChC,MAAU,MACN,wBAAwB,EAAI,OAAO,uCAAuC,EAAQ,EAAS,EAAE,EACjG,EAEJ,IAAM,EAAM,IAAI,WAAW,EAAI,MAAM,EACrC,IAAK,IAAI,EAAI,EAAG,EAAI,EAAI,OAAQ,GAAK,EACjC,EAAI,GAAK,EAAI,EAAI,GACjB,EAAI,EAAI,GAAK,EAAI,EAAI,GACrB,EAAI,EAAI,GAAK,EAAI,GAErB,OAAO,IAAI,EAAA,SAAS,EAAK,EAAO,CAAM,CAC1C,CAOA,SAAgB,EAAM,EAA6B,CAC/C,IAAM,EAAM,EAAM,KACZ,EAAM,IAAI,WAAW,EAAI,MAAM,EACrC,IAAK,IAAI,EAAI,EAAG,EAAI,EAAI,OAAQ,GAAK,EACjC,EAAI,GAAK,EAAI,EAAI,GACjB,EAAI,EAAI,GAAK,EAAI,EAAI,GACrB,EAAI,EAAI,GAAK,EAAI,GAErB,OAAO,CACX,CAoBA,SAAgB,EACZ,EACA,EACA,EACA,EAA0C,CAAC,IAAK,IAAK,GAAG,EACzC,CACf,IAAM,EAAQ,KAAK,IAAI,EAAc,EAAM,MAAO,EAAe,EAAM,MAAM,EACvE,EAAO,KAAK,MAAM,EAAM,MAAQ,CAAK,EACrC,EAAO,KAAK,MAAM,EAAM,OAAS,CAAK,EACtC,EAAU,EAAO,EAAO,EAAM,CAAI,EAElC,EAAM,IAAI,WAAW,EAAc,EAAe,CAAC,EACnD,EAAK,EAAK,GACV,EAAK,EAAK,GACV,EAAK,EAAK,GAChB,IAAK,IAAI,EAAI,EAAG,EAAI,EAAI,OAAQ,GAAK,EACjC,EAAI,GAAK,EACT,EAAI,EAAI,GAAK,EACb,EAAI,EAAI,GAAK,EAGjB,IAAM,EAAU,KAAK,OAAO,EAAc,GAAQ,CAAC,EAC7C,EAAS,KAAK,OAAO,EAAe,GAAQ,CAAC,EAC7C,EAAW,EAAO,EACxB,IAAK,IAAI,EAAI,EAAG,EAAI,EAAM,IAAK,CAC3B,IAAM,EAAY,EAAI,EAChB,IAAc,EAAS,GAAK,EAAc,GAAW,EAC3D,EAAI,IAAI,EAAQ,KAAK,SAAS,EAAW,EAAY,CAAQ,EAAG,CAAS,CAC7E,CAEA,MAAO,CACH,MAAO,IAAI,EAAA,SAAS,EAAK,EAAa,CAAY,EAClD,QACA,UACA,QACJ,CACJ"}

package/dist/vision/preprocess/image.js

@@ -1,93 +1,72 @@
-import * as v from "onnxruntime-web";
-import { createCanvas as y, get2DContext as h, rgbToImageData as I, imageDataToRGB as A } from "../core/canvas.js";
-import { RGBImage as $ } from "../types.js";
-function D(n, o, t) {
-  if (o <= 0 || t <= 0)
-    throw new Error(`Invalid resize target ${o}x${t}.`);
-  if (o === n.width && t === n.height)
-    return n;
-  const r = y(n.width, n.height);
-  h(r).putImageData(I(n), 0, 0);
-  const e = y(o, t), a = h(e);
-  a.imageSmoothingEnabled = !0, a.imageSmoothingQuality = "high", a.drawImage(r, 0, 0, o, t);
-  const l = a.getImageData(0, 0, o, t);
-  return A(l);
+import { RGBImage as e } from "../types.js";
+import { createCanvas as t, get2DContext as n, imageDataToRGB as r, rgbToImageData as i } from "../core/canvas.js";
+import * as a from "onnxruntime-web";
+//#region src/vision/preprocess/image.ts
+function o(e, a, o) {
+	if (a <= 0 || o <= 0) throw Error(`Invalid resize target ${a}x${o}.`);
+	if (a === e.width && o === e.height) return e;
+	let s = t(e.width, e.height);
+	n(s).putImageData(i(e), 0, 0);
+	let c = n(t(a, o));
+	return c.imageSmoothingEnabled = !0, c.imageSmoothingQuality = "high", c.drawImage(s, 0, 0, a, o), r(c.getImageData(0, 0, a, o));
 }
-function B(n, o, t, r = 1 / 255) {
-  const s = new Float32Array(n.data.length), e = n.data, a = o[0], l = o[1], c = o[2], x = t[0], d = t[1], w = t[2];
-  for (let f = 0; f < e.length; f += 3)
-    s[f] = (e[f] * r - a) / x, s[f + 1] = (e[f + 1] * r - l) / d, s[f + 2] = (e[f + 2] * r - c) / w;
-  return s;
+function s(e, t, n, r = 1 / 255) {
+	let i = new Float32Array(e.data.length), a = e.data, o = t[0], s = t[1], c = t[2], l = n[0], u = n[1], d = n[2];
+	for (let e = 0; e < a.length; e += 3) i[e] = (a[e] * r - o) / l, i[e + 1] = (a[e + 1] * r - s) / u, i[e + 2] = (a[e + 2] * r - c) / d;
+	return i;
 }
-function T(n, o = 1 / 255) {
-  const t = new Float32Array(n.data.length), r = n.data;
-  for (let s = 0; s < r.length; s++)
-    t[s] = r[s] * o;
-  return t;
+function c(e, t = 1 / 255) {
+	let n = new Float32Array(e.data.length), r = e.data;
+	for (let e = 0; e < r.length; e++) n[e] = r[e] * t;
+	return n;
 }
-function F(n, o, t, r = 3) {
-  const s = o * t * r;
-  if (n.length !== s)
-    throw new Error(
-      `toCHW: expected length ${s} for ${o}x${t}x${r}, got ${n.length}.`
-    );
-  const e = new Float32Array(s), a = o * t;
-  for (let l = 0; l < t; l++)
-    for (let c = 0; c < o; c++) {
-      const x = (l * o + c) * r, d = l * o + c;
-      for (let w = 0; w < r; w++)
-        e[w * a + d] = n[x + w];
-    }
-  return e;
+function l(e, t, n, r = 3) {
+	let i = t * n * r;
+	if (e.length !== i) throw Error(`toCHW: expected length ${i} for ${t}x${n}x${r}, got ${e.length}.`);
+	let a = new Float32Array(i), o = t * n;
+	for (let i = 0; i < n; i++) for (let n = 0; n < t; n++) {
+		let s = (i * t + n) * r, c = i * t + n;
+		for (let t = 0; t < r; t++) a[t * o + c] = e[s + t];
+	}
+	return a;
 }
-function E(n, o) {
-  return new v.Tensor("float32", n, o);
+function u(e, t) {
+	return new a.Tensor("float32", e, t);
 }
-function R(n) {
-  const o = T(n);
-  return F(o, n.width, n.height, 3);
+function d(e) {
+	return l(c(e), e.width, e.height, 3);
 }
-function U(n, o, t) {
-  if (n.length !== o * t * 3)
-    throw new Error(
-      `fromCv2: data length ${n.length} does not match width * height * 3 = ${o * t * 3}.`
-    );
-  const r = new Uint8Array(n.length);
-  for (let s = 0; s < n.length; s += 3)
-    r[s] = n[s + 2], r[s + 1] = n[s + 1], r[s + 2] = n[s];
-  return new $(r, o, t);
+function f(t, n, r) {
+	if (t.length !== n * r * 3) throw Error(`fromCv2: data length ${t.length} does not match width * height * 3 = ${n * r * 3}.`);
+	let i = new Uint8Array(t.length);
+	for (let e = 0; e < t.length; e += 3) i[e] = t[e + 2], i[e + 1] = t[e + 1], i[e + 2] = t[e];
+	return new e(i, n, r);
 }
-function G(n) {
-  const o = n.data, t = new Uint8Array(o.length);
-  for (let r = 0; r < o.length; r += 3)
-    t[r] = o[r + 2], t[r + 1] = o[r + 1], t[r + 2] = o[r];
-  return t;
+function p(e) {
+	let t = e.data, n = new Uint8Array(t.length);
+	for (let e = 0; e < t.length; e += 3) n[e] = t[e + 2], n[e + 1] = t[e + 1], n[e + 2] = t[e];
+	return n;
 }
-function O(n, o, t, r = [114, 114, 114]) {
-  const s = Math.min(o / n.width, t / n.height), e = Math.round(n.width * s), a = Math.round(n.height * s), l = D(n, e, a), c = new Uint8Array(o * t * 3), x = r[0], d = r[1], w = r[2];
-  for (let u = 0; u < c.length; u += 3)
-    c[u] = x, c[u + 1] = d, c[u + 2] = w;
-  const f = Math.floor((o - e) / 2), m = Math.floor((t - a) / 2), p = e * 3;
-  for (let u = 0; u < a; u++) {
-    const C = u * p, i = ((m + u) * o + f) * 3;
-    c.set(l.data.subarray(C, C + p), i);
-  }
-  return {
-    image: new $(c, o, t),
-    scale: s,
-    padLeft: f,
-    padTop: m
-  };
+function m(t, n, r, i = [
+	114,
+	114,
+	114
+]) {
+	let a = Math.min(n / t.width, r / t.height), s = Math.round(t.width * a), c = Math.round(t.height * a), l = o(t, s, c), u = new Uint8Array(n * r * 3), d = i[0], f = i[1], p = i[2];
+	for (let e = 0; e < u.length; e += 3) u[e] = d, u[e + 1] = f, u[e + 2] = p;
+	let m = Math.floor((n - s) / 2), h = Math.floor((r - c) / 2), g = s * 3;
+	for (let e = 0; e < c; e++) {
+		let t = e * g, r = ((h + e) * n + m) * 3;
+		u.set(l.data.subarray(t, t + g), r);
+	}
+	return {
+		image: new e(u, n, r),
+		scale: a,
+		padLeft: m,
+		padTop: h
+	};
 }
-export {
-  U as fromCv2,
-  O as letterbox,
-  B as normalize,
-  D as resize,
-  F as toCHW,
-  G as toCv2,
-  T as toFloat32,
-  E as toFloat32Tensor,
-  R as toTensor
-};
-//# sourceMappingURL=image.js.map
+//#endregion
+export { f as fromCv2, m as letterbox, s as normalize, o as resize, l as toCHW, p as toCv2, c as toFloat32, u as toFloat32Tensor, d as toTensor };
+
+//# sourceMappingURL=image.js.map

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

@@ -1 +1 @@
-{"version":3,"file":"image.js","sources":["../../../src/vision/preprocess/image.ts"],"sourcesContent":["/**\n * Composable image preprocessing primitives (resize, normalize, letterbox, layout).\n *\n * Mirrors the Python `preprocess.image` module. Operates on the canonical\n * {@link RGBImage} (HWC RGB uint8) and produces uint8 or float32 buffers\n * depending on the operation.\n */\n\nimport * as ortRuntime from \"onnxruntime-web\";\nimport type * as ort from \"onnxruntime-web\";\n\nimport { createCanvas, get2DContext, imageDataToRGB, rgbToImageData } from \"../core/canvas\";\nimport { RGBImage } from \"../types\";\n\n/** Resize an image to `(targetWidth, targetHeight)` using high-quality canvas resampling. */\nexport function resize(image: RGBImage, targetWidth: number, targetHeight: number): RGBImage {\n    if (targetWidth <= 0 || targetHeight <= 0) {\n        throw new Error(`Invalid resize target ${targetWidth}x${targetHeight}.`);\n    }\n    if (targetWidth === image.width && targetHeight === image.height) {\n        return image;\n    }\n\n    const srcCanvas = createCanvas(image.width, image.height);\n    const srcCtx = get2DContext(srcCanvas);\n    srcCtx.putImageData(rgbToImageData(image), 0, 0);\n\n    const dstCanvas = createCanvas(targetWidth, targetHeight);\n    const dstCtx = get2DContext(dstCanvas);\n    dstCtx.imageSmoothingEnabled = true;\n    dstCtx.imageSmoothingQuality = \"high\";\n    dstCtx.drawImage(srcCanvas as CanvasImageSource, 0, 0, targetWidth, targetHeight);\n\n    const imageData = dstCtx.getImageData(0, 0, targetWidth, targetHeight);\n    return imageDataToRGB(imageData);\n}\n\n/**\n * Convert a uint8 image to a normalized float32 array (HWC layout preserved).\n *\n * Applies `(pixel * scale - mean) / std` channel-wise.\n */\nexport function normalize(\n    image: RGBImage,\n    mean: readonly [number, number, number],\n    std: readonly [number, number, number],\n    scale: number = 1 / 255,\n): Float32Array {\n    const out = new Float32Array(image.data.length);\n    const data = image.data;\n    const m0 = mean[0];\n    const m1 = mean[1];\n    const m2 = mean[2];\n    const s0 = std[0];\n    const s1 = std[1];\n    const s2 = std[2];\n    for (let i = 0; i < data.length; i += 3) {\n        out[i] = ((data[i] as number) * scale - m0) / s0;\n        out[i + 1] = ((data[i + 1] as number) * scale - m1) / s1;\n        out[i + 2] = ((data[i + 2] as number) * scale - m2) / s2;\n    }\n    return out;\n}\n\n/** Convert a uint8 image to a `Float32Array` in `[0, 1]` (HWC layout preserved). */\nexport function toFloat32(image: RGBImage, scale: number = 1 / 255): Float32Array {\n    const out = new Float32Array(image.data.length);\n    const data = image.data;\n    for (let i = 0; i < data.length; i++) {\n        out[i] = (data[i] as number) * scale;\n    }\n    return out;\n}\n\n/**\n * Transpose interleaved HWC data to planar CHW layout.\n *\n * @param hwc Source array of length `width * height * channels`.\n */\nexport function toCHW(\n    hwc: Float32Array,\n    width: number,\n    height: number,\n    channels: number = 3,\n): Float32Array {\n    const expected = width * height * channels;\n    if (hwc.length !== expected) {\n        throw new Error(\n            `toCHW: expected length ${expected} for ${width}x${height}x${channels}, got ${hwc.length}.`,\n        );\n    }\n    const chw = new Float32Array(expected);\n    const plane = width * height;\n    for (let y = 0; y < height; y++) {\n        for (let x = 0; x < width; x++) {\n            const hwcBase = (y * width + x) * channels;\n            const planeIdx = y * width + x;\n            for (let c = 0; c < channels; c++) {\n                chw[c * plane + planeIdx] = hwc[hwcBase + c] as number;\n            }\n        }\n    }\n    return chw;\n}\n\n/** Wrap a Float32 buffer into an `ort.Tensor`. */\nexport function toFloat32Tensor(data: Float32Array, dims: readonly number[]): ort.Tensor {\n    return new ortRuntime.Tensor(\"float32\", data, dims as number[]);\n}\n\n/**\n * Convert an HWC uint8 image to a CHW `Float32Array` scaled to `[0, 1]`.\n *\n * Mirrors `torchvision.transforms.ToTensor()` semantics: HWC → CHW,\n * `uint8 → float32 / 255`. Useful as input to YOLO-style detectors that\n * don't require ImageNet normalization.\n *\n * @returns CHW `Float32Array` of length `width * height * 3`.\n */\nexport function toTensor(image: RGBImage): Float32Array {\n    const f32 = toFloat32(image);\n    return toCHW(f32, image.width, image.height, 3);\n}\n\n/**\n * Convert an HWC BGR uint8 buffer (OpenCV layout) to the SDK's HWC RGB.\n *\n * Use when you receive image bytes from `cv2.imencode` over the wire and\n * want to feed them to the SDK without going through a canvas decode.\n *\n * @param bgr Flat BGR Uint8Array of length `width * height * 3`.\n */\nexport function fromCv2(bgr: Uint8Array, width: number, height: number): RGBImage {\n    if (bgr.length !== width * height * 3) {\n        throw new Error(\n            `fromCv2: data length ${bgr.length} does not match width * height * 3 = ${width * height * 3}.`,\n        );\n    }\n    const rgb = new Uint8Array(bgr.length);\n    for (let i = 0; i < bgr.length; i += 3) {\n        rgb[i] = bgr[i + 2] as number;\n        rgb[i + 1] = bgr[i + 1] as number;\n        rgb[i + 2] = bgr[i] as number;\n    }\n    return new RGBImage(rgb, width, height);\n}\n\n/**\n * Convert the SDK's HWC RGB image to an HWC BGR `Uint8Array` (OpenCV layout).\n *\n * Useful for round-tripping data to a Python OpenCV consumer.\n */\nexport function toCv2(image: RGBImage): Uint8Array {\n    const rgb = image.data;\n    const bgr = new Uint8Array(rgb.length);\n    for (let i = 0; i < rgb.length; i += 3) {\n        bgr[i] = rgb[i + 2] as number;\n        bgr[i + 1] = rgb[i + 1] as number;\n        bgr[i + 2] = rgb[i] as number;\n    }\n    return bgr;\n}\n\nexport interface LetterboxResult {\n    /** The padded image at the target size. */\n    readonly image: RGBImage;\n    /** The factor applied to the original image (`< 1` if downscaled). */\n    readonly scale: number;\n    /** Horizontal padding in pixels (left side; right side has the same or +1). */\n    readonly padLeft: number;\n    /** Vertical padding in pixels (top side). */\n    readonly padTop: number;\n}\n\n/**\n * Resize preserving aspect ratio, padding to `(targetWidth, targetHeight)`\n * with a constant fill color.\n *\n * Standard YOLO preprocessing — returning `scale` and `padLeft`/`padTop`\n * lets callers map detections back to the original image coordinates.\n */\nexport function letterbox(\n    image: RGBImage,\n    targetWidth: number,\n    targetHeight: number,\n    fill: readonly [number, number, number] = [114, 114, 114],\n): LetterboxResult {\n    const scale = Math.min(targetWidth / image.width, targetHeight / image.height);\n    const newW = Math.round(image.width * scale);\n    const newH = Math.round(image.height * scale);\n    const resized = resize(image, newW, newH);\n\n    const out = new Uint8Array(targetWidth * targetHeight * 3);\n    const f0 = fill[0];\n    const f1 = fill[1];\n    const f2 = fill[2];\n    for (let i = 0; i < out.length; i += 3) {\n        out[i] = f0;\n        out[i + 1] = f1;\n        out[i + 2] = f2;\n    }\n\n    const padLeft = Math.floor((targetWidth - newW) / 2);\n    const padTop = Math.floor((targetHeight - newH) / 2);\n    const rowBytes = newW * 3;\n    for (let y = 0; y < newH; y++) {\n        const srcOffset = y * rowBytes;\n        const dstOffset = ((padTop + y) * targetWidth + padLeft) * 3;\n        out.set(resized.data.subarray(srcOffset, srcOffset + rowBytes), dstOffset);\n    }\n\n    return {\n        image: new RGBImage(out, targetWidth, targetHeight),\n        scale,\n        padLeft,\n        padTop,\n    };\n}\n"],"names":["resize","image","targetWidth","targetHeight","srcCanvas","createCanvas","get2DContext","rgbToImageData","dstCanvas","dstCtx","imageData","imageDataToRGB","normalize","mean","std","scale","out","data","m0","m1","m2","s0","s1","s2","i","toFloat32","toCHW","hwc","width","height","channels","expected","chw","plane","y","x","hwcBase","planeIdx","c","toFloat32Tensor","dims","ortRuntime","toTensor","f32","fromCv2","bgr","rgb","RGBImage","toCv2","letterbox","fill","newW","newH","resized","f0","f1","f2","padLeft","padTop","rowBytes","srcOffset","dstOffset"],"mappings":";;;AAeO,SAASA,EAAOC,GAAiBC,GAAqBC,GAAgC;AACzF,MAAID,KAAe,KAAKC,KAAgB;AACpC,UAAM,IAAI,MAAM,yBAAyBD,CAAW,IAAIC,CAAY,GAAG;AAE3E,MAAID,MAAgBD,EAAM,SAASE,MAAiBF,EAAM;AACtD,WAAOA;AAGX,QAAMG,IAAYC,EAAaJ,EAAM,OAAOA,EAAM,MAAM;AAExD,EADeK,EAAaF,CAAS,EAC9B,aAAaG,EAAeN,CAAK,GAAG,GAAG,CAAC;AAE/C,QAAMO,IAAYH,EAAaH,GAAaC,CAAY,GAClDM,IAASH,EAAaE,CAAS;AACrC,EAAAC,EAAO,wBAAwB,IAC/BA,EAAO,wBAAwB,QAC/BA,EAAO,UAAUL,GAAgC,GAAG,GAAGF,GAAaC,CAAY;AAEhF,QAAMO,IAAYD,EAAO,aAAa,GAAG,GAAGP,GAAaC,CAAY;AACrE,SAAOQ,EAAeD,CAAS;AACnC;AAOO,SAASE,EACZX,GACAY,GACAC,GACAC,IAAgB,IAAI,KACR;AACZ,QAAMC,IAAM,IAAI,aAAaf,EAAM,KAAK,MAAM,GACxCgB,IAAOhB,EAAM,MACbiB,IAAKL,EAAK,CAAC,GACXM,IAAKN,EAAK,CAAC,GACXO,IAAKP,EAAK,CAAC,GACXQ,IAAKP,EAAI,CAAC,GACVQ,IAAKR,EAAI,CAAC,GACVS,IAAKT,EAAI,CAAC;AAChB,WAASU,IAAI,GAAGA,IAAIP,EAAK,QAAQO,KAAK;AAClC,IAAAR,EAAIQ,CAAC,KAAMP,EAAKO,CAAC,IAAeT,IAAQG,KAAMG,GAC9CL,EAAIQ,IAAI,CAAC,KAAMP,EAAKO,IAAI,CAAC,IAAeT,IAAQI,KAAMG,GACtDN,EAAIQ,IAAI,CAAC,KAAMP,EAAKO,IAAI,CAAC,IAAeT,IAAQK,KAAMG;AAE1D,SAAOP;AACX;AAGO,SAASS,EAAUxB,GAAiBc,IAAgB,IAAI,KAAmB;AAC9E,QAAMC,IAAM,IAAI,aAAaf,EAAM,KAAK,MAAM,GACxCgB,IAAOhB,EAAM;AACnB,WAASuB,IAAI,GAAGA,IAAIP,EAAK,QAAQO;AAC7B,IAAAR,EAAIQ,CAAC,IAAKP,EAAKO,CAAC,IAAeT;AAEnC,SAAOC;AACX;AAOO,SAASU,EACZC,GACAC,GACAC,GACAC,IAAmB,GACP;AACZ,QAAMC,IAAWH,IAAQC,IAASC;AAClC,MAAIH,EAAI,WAAWI;AACf,UAAM,IAAI;AAAA,MACN,0BAA0BA,CAAQ,QAAQH,CAAK,IAAIC,CAAM,IAAIC,CAAQ,SAASH,EAAI,MAAM;AAAA,IAAA;AAGhG,QAAMK,IAAM,IAAI,aAAaD,CAAQ,GAC/BE,IAAQL,IAAQC;AACtB,WAASK,IAAI,GAAGA,IAAIL,GAAQK;AACxB,aAASC,IAAI,GAAGA,IAAIP,GAAOO,KAAK;AAC5B,YAAMC,KAAWF,IAAIN,IAAQO,KAAKL,GAC5BO,IAAWH,IAAIN,IAAQO;AAC7B,eAASG,IAAI,GAAGA,IAAIR,GAAUQ;AAC1B,QAAAN,EAAIM,IAAIL,IAAQI,CAAQ,IAAIV,EAAIS,IAAUE,CAAC;AAAA,IAEnD;AAEJ,SAAON;AACX;AAGO,SAASO,EAAgBtB,GAAoBuB,GAAqC;AACrF,SAAO,IAAIC,EAAW,OAAO,WAAWxB,GAAMuB,CAAgB;AAClE;AAWO,SAASE,EAASzC,GAA+B;AACpD,QAAM0C,IAAMlB,EAAUxB,CAAK;AAC3B,SAAOyB,EAAMiB,GAAK1C,EAAM,OAAOA,EAAM,QAAQ,CAAC;AAClD;AAUO,SAAS2C,EAAQC,GAAiBjB,GAAeC,GAA0B;AAC9E,MAAIgB,EAAI,WAAWjB,IAAQC,IAAS;AAChC,UAAM,IAAI;AAAA,MACN,wBAAwBgB,EAAI,MAAM,wCAAwCjB,IAAQC,IAAS,CAAC;AAAA,IAAA;AAGpG,QAAMiB,IAAM,IAAI,WAAWD,EAAI,MAAM;AACrC,WAASrB,IAAI,GAAGA,IAAIqB,EAAI,QAAQrB,KAAK;AACjC,IAAAsB,EAAItB,CAAC,IAAIqB,EAAIrB,IAAI,CAAC,GAClBsB,EAAItB,IAAI,CAAC,IAAIqB,EAAIrB,IAAI,CAAC,GACtBsB,EAAItB,IAAI,CAAC,IAAIqB,EAAIrB,CAAC;AAEtB,SAAO,IAAIuB,EAASD,GAAKlB,GAAOC,CAAM;AAC1C;AAOO,SAASmB,EAAM/C,GAA6B;AAC/C,QAAM6C,IAAM7C,EAAM,MACZ4C,IAAM,IAAI,WAAWC,EAAI,MAAM;AACrC,WAAStB,IAAI,GAAGA,IAAIsB,EAAI,QAAQtB,KAAK;AACjC,IAAAqB,EAAIrB,CAAC,IAAIsB,EAAItB,IAAI,CAAC,GAClBqB,EAAIrB,IAAI,CAAC,IAAIsB,EAAItB,IAAI,CAAC,GACtBqB,EAAIrB,IAAI,CAAC,IAAIsB,EAAItB,CAAC;AAEtB,SAAOqB;AACX;AAoBO,SAASI,EACZhD,GACAC,GACAC,GACA+C,IAA0C,CAAC,KAAK,KAAK,GAAG,GACzC;AACf,QAAMnC,IAAQ,KAAK,IAAIb,IAAcD,EAAM,OAAOE,IAAeF,EAAM,MAAM,GACvEkD,IAAO,KAAK,MAAMlD,EAAM,QAAQc,CAAK,GACrCqC,IAAO,KAAK,MAAMnD,EAAM,SAASc,CAAK,GACtCsC,IAAUrD,EAAOC,GAAOkD,GAAMC,CAAI,GAElCpC,IAAM,IAAI,WAAWd,IAAcC,IAAe,CAAC,GACnDmD,IAAKJ,EAAK,CAAC,GACXK,IAAKL,EAAK,CAAC,GACXM,IAAKN,EAAK,CAAC;AACjB,WAAS1B,IAAI,GAAGA,IAAIR,EAAI,QAAQQ,KAAK;AACjC,IAAAR,EAAIQ,CAAC,IAAI8B,GACTtC,EAAIQ,IAAI,CAAC,IAAI+B,GACbvC,EAAIQ,IAAI,CAAC,IAAIgC;AAGjB,QAAMC,IAAU,KAAK,OAAOvD,IAAciD,KAAQ,CAAC,GAC7CO,IAAS,KAAK,OAAOvD,IAAeiD,KAAQ,CAAC,GAC7CO,IAAWR,IAAO;AACxB,WAASjB,IAAI,GAAGA,IAAIkB,GAAMlB,KAAK;AAC3B,UAAM0B,IAAY1B,IAAIyB,GAChBE,MAAcH,IAASxB,KAAKhC,IAAcuD,KAAW;AAC3D,IAAAzC,EAAI,IAAIqC,EAAQ,KAAK,SAASO,GAAWA,IAAYD,CAAQ,GAAGE,CAAS;AAAA,EAC7E;AAEA,SAAO;AAAA,IACH,OAAO,IAAId,EAAS/B,GAAKd,GAAaC,CAAY;AAAA,IAClD,OAAAY;AAAA,IACA,SAAA0C;AAAA,IACA,QAAAC;AAAA,EAAA;AAER;"}
+{"version":3,"file":"image.js","names":[],"sources":["../../../src/vision/preprocess/image.ts"],"sourcesContent":["/**\n * Composable image preprocessing primitives (resize, normalize, letterbox, layout).\n *\n * Mirrors the Python `preprocess.image` module. Operates on the canonical\n * {@link RGBImage} (HWC RGB uint8) and produces uint8 or float32 buffers\n * depending on the operation.\n */\n\nimport * as ortRuntime from \"onnxruntime-web\";\nimport type * as ort from \"onnxruntime-web\";\n\nimport { createCanvas, get2DContext, imageDataToRGB, rgbToImageData } from \"../core/canvas\";\nimport { RGBImage } from \"../types\";\n\n/** Resize an image to `(targetWidth, targetHeight)` using high-quality canvas resampling. */\nexport function resize(image: RGBImage, targetWidth: number, targetHeight: number): RGBImage {\n    if (targetWidth <= 0 || targetHeight <= 0) {\n        throw new Error(`Invalid resize target ${targetWidth}x${targetHeight}.`);\n    }\n    if (targetWidth === image.width && targetHeight === image.height) {\n        return image;\n    }\n\n    const srcCanvas = createCanvas(image.width, image.height);\n    const srcCtx = get2DContext(srcCanvas);\n    srcCtx.putImageData(rgbToImageData(image), 0, 0);\n\n    const dstCanvas = createCanvas(targetWidth, targetHeight);\n    const dstCtx = get2DContext(dstCanvas);\n    dstCtx.imageSmoothingEnabled = true;\n    dstCtx.imageSmoothingQuality = \"high\";\n    dstCtx.drawImage(srcCanvas as CanvasImageSource, 0, 0, targetWidth, targetHeight);\n\n    const imageData = dstCtx.getImageData(0, 0, targetWidth, targetHeight);\n    return imageDataToRGB(imageData);\n}\n\n/**\n * Convert a uint8 image to a normalized float32 array (HWC layout preserved).\n *\n * Applies `(pixel * scale - mean) / std` channel-wise.\n */\nexport function normalize(\n    image: RGBImage,\n    mean: readonly [number, number, number],\n    std: readonly [number, number, number],\n    scale: number = 1 / 255,\n): Float32Array {\n    const out = new Float32Array(image.data.length);\n    const data = image.data;\n    const m0 = mean[0];\n    const m1 = mean[1];\n    const m2 = mean[2];\n    const s0 = std[0];\n    const s1 = std[1];\n    const s2 = std[2];\n    for (let i = 0; i < data.length; i += 3) {\n        out[i] = ((data[i] as number) * scale - m0) / s0;\n        out[i + 1] = ((data[i + 1] as number) * scale - m1) / s1;\n        out[i + 2] = ((data[i + 2] as number) * scale - m2) / s2;\n    }\n    return out;\n}\n\n/** Convert a uint8 image to a `Float32Array` in `[0, 1]` (HWC layout preserved). */\nexport function toFloat32(image: RGBImage, scale: number = 1 / 255): Float32Array {\n    const out = new Float32Array(image.data.length);\n    const data = image.data;\n    for (let i = 0; i < data.length; i++) {\n        out[i] = (data[i] as number) * scale;\n    }\n    return out;\n}\n\n/**\n * Transpose interleaved HWC data to planar CHW layout.\n *\n * @param hwc Source array of length `width * height * channels`.\n */\nexport function toCHW(\n    hwc: Float32Array,\n    width: number,\n    height: number,\n    channels: number = 3,\n): Float32Array {\n    const expected = width * height * channels;\n    if (hwc.length !== expected) {\n        throw new Error(\n            `toCHW: expected length ${expected} for ${width}x${height}x${channels}, got ${hwc.length}.`,\n        );\n    }\n    const chw = new Float32Array(expected);\n    const plane = width * height;\n    for (let y = 0; y < height; y++) {\n        for (let x = 0; x < width; x++) {\n            const hwcBase = (y * width + x) * channels;\n            const planeIdx = y * width + x;\n            for (let c = 0; c < channels; c++) {\n                chw[c * plane + planeIdx] = hwc[hwcBase + c] as number;\n            }\n        }\n    }\n    return chw;\n}\n\n/** Wrap a Float32 buffer into an `ort.Tensor`. */\nexport function toFloat32Tensor(data: Float32Array, dims: readonly number[]): ort.Tensor {\n    return new ortRuntime.Tensor(\"float32\", data, dims as number[]);\n}\n\n/**\n * Convert an HWC uint8 image to a CHW `Float32Array` scaled to `[0, 1]`.\n *\n * Mirrors `torchvision.transforms.ToTensor()` semantics: HWC → CHW,\n * `uint8 → float32 / 255`. Useful as input to YOLO-style detectors that\n * don't require ImageNet normalization.\n *\n * @returns CHW `Float32Array` of length `width * height * 3`.\n */\nexport function toTensor(image: RGBImage): Float32Array {\n    const f32 = toFloat32(image);\n    return toCHW(f32, image.width, image.height, 3);\n}\n\n/**\n * Convert an HWC BGR uint8 buffer (OpenCV layout) to the SDK's HWC RGB.\n *\n * Use when you receive image bytes from `cv2.imencode` over the wire and\n * want to feed them to the SDK without going through a canvas decode.\n *\n * @param bgr Flat BGR Uint8Array of length `width * height * 3`.\n */\nexport function fromCv2(bgr: Uint8Array, width: number, height: number): RGBImage {\n    if (bgr.length !== width * height * 3) {\n        throw new Error(\n            `fromCv2: data length ${bgr.length} does not match width * height * 3 = ${width * height * 3}.`,\n        );\n    }\n    const rgb = new Uint8Array(bgr.length);\n    for (let i = 0; i < bgr.length; i += 3) {\n        rgb[i] = bgr[i + 2] as number;\n        rgb[i + 1] = bgr[i + 1] as number;\n        rgb[i + 2] = bgr[i] as number;\n    }\n    return new RGBImage(rgb, width, height);\n}\n\n/**\n * Convert the SDK's HWC RGB image to an HWC BGR `Uint8Array` (OpenCV layout).\n *\n * Useful for round-tripping data to a Python OpenCV consumer.\n */\nexport function toCv2(image: RGBImage): Uint8Array {\n    const rgb = image.data;\n    const bgr = new Uint8Array(rgb.length);\n    for (let i = 0; i < rgb.length; i += 3) {\n        bgr[i] = rgb[i + 2] as number;\n        bgr[i + 1] = rgb[i + 1] as number;\n        bgr[i + 2] = rgb[i] as number;\n    }\n    return bgr;\n}\n\nexport interface LetterboxResult {\n    /** The padded image at the target size. */\n    readonly image: RGBImage;\n    /** The factor applied to the original image (`< 1` if downscaled). */\n    readonly scale: number;\n    /** Horizontal padding in pixels (left side; right side has the same or +1). */\n    readonly padLeft: number;\n    /** Vertical padding in pixels (top side). */\n    readonly padTop: number;\n}\n\n/**\n * Resize preserving aspect ratio, padding to `(targetWidth, targetHeight)`\n * with a constant fill color.\n *\n * Standard YOLO preprocessing — returning `scale` and `padLeft`/`padTop`\n * lets callers map detections back to the original image coordinates.\n */\nexport function letterbox(\n    image: RGBImage,\n    targetWidth: number,\n    targetHeight: number,\n    fill: readonly [number, number, number] = [114, 114, 114],\n): LetterboxResult {\n    const scale = Math.min(targetWidth / image.width, targetHeight / image.height);\n    const newW = Math.round(image.width * scale);\n    const newH = Math.round(image.height * scale);\n    const resized = resize(image, newW, newH);\n\n    const out = new Uint8Array(targetWidth * targetHeight * 3);\n    const f0 = fill[0];\n    const f1 = fill[1];\n    const f2 = fill[2];\n    for (let i = 0; i < out.length; i += 3) {\n        out[i] = f0;\n        out[i + 1] = f1;\n        out[i + 2] = f2;\n    }\n\n    const padLeft = Math.floor((targetWidth - newW) / 2);\n    const padTop = Math.floor((targetHeight - newH) / 2);\n    const rowBytes = newW * 3;\n    for (let y = 0; y < newH; y++) {\n        const srcOffset = y * rowBytes;\n        const dstOffset = ((padTop + y) * targetWidth + padLeft) * 3;\n        out.set(resized.data.subarray(srcOffset, srcOffset + rowBytes), dstOffset);\n    }\n\n    return {\n        image: new RGBImage(out, targetWidth, targetHeight),\n        scale,\n        padLeft,\n        padTop,\n    };\n}\n"],"mappings":";;;;AAeA,SAAgB,EAAO,GAAiB,GAAqB,GAAgC;CACzF,IAAI,KAAe,KAAK,KAAgB,GACpC,MAAU,MAAM,yBAAyB,EAAY,GAAG,EAAa,EAAE;CAE3E,IAAI,MAAgB,EAAM,SAAS,MAAiB,EAAM,QACtD,OAAO;CAGX,IAAM,IAAY,EAAa,EAAM,OAAO,EAAM,MAAM;CAExD,EAD4B,CAC5B,CAAA,CAAO,aAAa,EAAe,CAAK,GAAG,GAAG,CAAC;CAG/C,IAAM,IAAS,EADG,EAAa,GAAa,CAChB,CAAS;CAMrC,OALA,EAAO,wBAAwB,IAC/B,EAAO,wBAAwB,QAC/B,EAAO,UAAU,GAAgC,GAAG,GAAG,GAAa,CAAY,GAGzE,EADW,EAAO,aAAa,GAAG,GAAG,GAAa,CACnC,CAAS;AACnC;AAOA,SAAgB,EACZ,GACA,GACA,GACA,IAAgB,IAAI,KACR;CACZ,IAAM,IAAM,IAAI,aAAa,EAAM,KAAK,MAAM,GACxC,IAAO,EAAM,MACb,IAAK,EAAK,IACV,IAAK,EAAK,IACV,IAAK,EAAK,IACV,IAAK,EAAI,IACT,IAAK,EAAI,IACT,IAAK,EAAI;CACf,KAAK,IAAI,IAAI,GAAG,IAAI,EAAK,QAAQ,KAAK,GAGlC,AAFA,EAAI,MAAO,EAAK,KAAgB,IAAQ,KAAM,GAC9C,EAAI,IAAI,MAAO,EAAK,IAAI,KAAgB,IAAQ,KAAM,GACtD,EAAI,IAAI,MAAO,EAAK,IAAI,KAAgB,IAAQ,KAAM;CAE1D,OAAO;AACX;AAGA,SAAgB,EAAU,GAAiB,IAAgB,IAAI,KAAmB;CAC9E,IAAM,IAAM,IAAI,aAAa,EAAM,KAAK,MAAM,GACxC,IAAO,EAAM;CACnB,KAAK,IAAI,IAAI,GAAG,IAAI,EAAK,QAAQ,KAC7B,EAAI,KAAM,EAAK,KAAgB;CAEnC,OAAO;AACX;AAOA,SAAgB,EACZ,GACA,GACA,GACA,IAAmB,GACP;CACZ,IAAM,IAAW,IAAQ,IAAS;CAClC,IAAI,EAAI,WAAW,GACf,MAAU,MACN,0BAA0B,EAAS,OAAO,EAAM,GAAG,EAAO,GAAG,EAAS,QAAQ,EAAI,OAAO,EAC7F;CAEJ,IAAM,IAAM,IAAI,aAAa,CAAQ,GAC/B,IAAQ,IAAQ;CACtB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAQ,KACxB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAO,KAAK;EAC5B,IAAM,KAAW,IAAI,IAAQ,KAAK,GAC5B,IAAW,IAAI,IAAQ;EAC7B,KAAK,IAAI,IAAI,GAAG,IAAI,GAAU,KAC1B,EAAI,IAAI,IAAQ,KAAY,EAAI,IAAU;CAElD;CAEJ,OAAO;AACX;AAGA,SAAgB,EAAgB,GAAoB,GAAqC;CACrF,OAAO,IAAI,EAAW,OAAO,WAAW,GAAM,CAAgB;AAClE;AAWA,SAAgB,EAAS,GAA+B;CAEpD,OAAO,EADK,EAAU,CACT,GAAK,EAAM,OAAO,EAAM,QAAQ,CAAC;AAClD;AAUA,SAAgB,EAAQ,GAAiB,GAAe,GAA0B;CAC9E,IAAI,EAAI,WAAW,IAAQ,IAAS,GAChC,MAAU,MACN,wBAAwB,EAAI,OAAO,uCAAuC,IAAQ,IAAS,EAAE,EACjG;CAEJ,IAAM,IAAM,IAAI,WAAW,EAAI,MAAM;CACrC,KAAK,IAAI,IAAI,GAAG,IAAI,EAAI,QAAQ,KAAK,GAGjC,AAFA,EAAI,KAAK,EAAI,IAAI,IACjB,EAAI,IAAI,KAAK,EAAI,IAAI,IACrB,EAAI,IAAI,KAAK,EAAI;CAErB,OAAO,IAAI,EAAS,GAAK,GAAO,CAAM;AAC1C;AAOA,SAAgB,EAAM,GAA6B;CAC/C,IAAM,IAAM,EAAM,MACZ,IAAM,IAAI,WAAW,EAAI,MAAM;CACrC,KAAK,IAAI,IAAI,GAAG,IAAI,EAAI,QAAQ,KAAK,GAGjC,AAFA,EAAI,KAAK,EAAI,IAAI,IACjB,EAAI,IAAI,KAAK,EAAI,IAAI,IACrB,EAAI,IAAI,KAAK,EAAI;CAErB,OAAO;AACX;AAoBA,SAAgB,EACZ,GACA,GACA,GACA,IAA0C;CAAC;CAAK;CAAK;AAAG,GACzC;CACf,IAAM,IAAQ,KAAK,IAAI,IAAc,EAAM,OAAO,IAAe,EAAM,MAAM,GACvE,IAAO,KAAK,MAAM,EAAM,QAAQ,CAAK,GACrC,IAAO,KAAK,MAAM,EAAM,SAAS,CAAK,GACtC,IAAU,EAAO,GAAO,GAAM,CAAI,GAElC,IAAM,IAAI,WAAW,IAAc,IAAe,CAAC,GACnD,IAAK,EAAK,IACV,IAAK,EAAK,IACV,IAAK,EAAK;CAChB,KAAK,IAAI,IAAI,GAAG,IAAI,EAAI,QAAQ,KAAK,GAGjC,AAFA,EAAI,KAAK,GACT,EAAI,IAAI,KAAK,GACb,EAAI,IAAI,KAAK;CAGjB,IAAM,IAAU,KAAK,OAAO,IAAc,KAAQ,CAAC,GAC7C,IAAS,KAAK,OAAO,IAAe,KAAQ,CAAC,GAC7C,IAAW,IAAO;CACxB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAM,KAAK;EAC3B,IAAM,IAAY,IAAI,GAChB,MAAc,IAAS,KAAK,IAAc,KAAW;EAC3D,EAAI,IAAI,EAAQ,KAAK,SAAS,GAAW,IAAY,CAAQ,GAAG,CAAS;CAC7E;CAEA,OAAO;EACH,OAAO,IAAI,EAAS,GAAK,GAAa,CAAY;EAClD;EACA;EACA;CACJ;AACJ"}