Spaces:
Sleeping
Sleeping
File size: 14,320 Bytes
e56fe87 a125618 e56fe87 a125618 e56fe87 a125618 e56fe87 a125618 e56fe87 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 | // Auto-segmentation using SlimSAM (Segment Anything Model)
// Generates a grid of point prompts, runs mask decoder for each,
// filters and deduplicates to find distinct parts of a drawing.
import { getTransformers } from './segmentation.js';
const SAM_MODEL = 'Xenova/slimsam-77-uniform';
const SAM_DIM = 384; // smaller input → quadratically less peak memory
const GRID_SIZE = 6; // 6x6 = 36 points, filtered to non-transparent
const MIN_IOU_SCORE = 0.65;
const MIN_AREA_FRAC = 0.005; // minimum mask area as fraction of image
const NMS_IOU_THRESHOLD = 0.5;
/**
* Auto-segment an image into distinct parts using SlimSAM.
* @param {Blob} imageBlob - background-removed PNG
* @param {function} onProgress - { message, progress }
* @returns {Promise<SegmentResult[]>}
*/
export async function autoSegment(imageBlob, onProgress) {
const { SamModel, AutoProcessor, RawImage } = await getTransformers();
// Prefer WebGPU (fast), fall back to WASM (slow but universal)
const hasWebGPU = typeof navigator !== 'undefined' && !!navigator.gpu;
const device = hasWebGPU ? 'webgpu' : 'wasm';
if (onProgress) onProgress({ message: `Loading segmentation model (${device})...`, progress: 0 });
const dtypeProgress = (p) => {
if (onProgress && p.progress != null) {
onProgress({
message: `Downloading model: ${Math.round(p.progress)}%`,
progress: p.progress * 0.2, // 0-20% for download
});
}
};
// Try the smallest dtype first; fall back step-by-step if the model
// doesn't ship that variant or the runtime can't load it.
const dtypePreference = device === 'webgpu'
? ['fp16', 'q8', 'fp32']
: ['q8', 'fp16', 'fp32'];
let model = null;
let loadedDtype = null;
let lastErr = null;
for (const dtype of dtypePreference) {
try {
model = await SamModel.from_pretrained(SAM_MODEL, {
device,
dtype,
progress_callback: dtypeProgress,
});
loadedDtype = dtype;
console.log(`[auto-segment] loaded SAM with dtype=${dtype}, device=${device}`);
break;
} catch (err) {
console.warn(`[auto-segment] dtype=${dtype} failed:`, err && err.message);
lastErr = err;
}
}
if (!model) throw lastErr || new Error('Failed to load SAM');
if (onProgress) onProgress({ message: `sam: loaded ${loadedDtype}/${device}`, progress: 22 });
const processor = await AutoProcessor.from_pretrained(SAM_MODEL);
if (onProgress) onProgress({ message: 'Preparing image...', progress: 25 });
// Downscale for inference
const workBlob = await downscale(imageBlob, SAM_DIM);
const workBitmap = await createImageBitmap(workBlob);
const imgW = workBitmap.width;
const imgH = workBitmap.height;
// Build opacity map to skip transparent grid points (bg-removed image)
const opaCanvas = new OffscreenCanvas(imgW, imgH);
const opaCtx = opaCanvas.getContext('2d');
opaCtx.drawImage(workBitmap, 0, 0);
const opaData = opaCtx.getImageData(0, 0, imgW, imgH).data;
workBitmap.close();
// Load as RawImage for the processor
const url = URL.createObjectURL(workBlob);
const rawImage = await RawImage.fromURL(url);
URL.revokeObjectURL(url);
if (onProgress) onProgress({ message: 'Analyzing image...', progress: 30 });
// Generate grid points, filter to non-transparent pixels
const allGridPoints = generateGrid(imgW, imgH, GRID_SIZE);
const gridPoints = allGridPoints.filter(([px, py]) => {
// Check a small area around the point for any opaque pixel
const r = 3;
for (let dy = -r; dy <= r; dy++) {
for (let dx = -r; dx <= r; dx++) {
const sx = Math.max(0, Math.min(imgW - 1, px + dx));
const sy = Math.max(0, Math.min(imgH - 1, py + dy));
if (opaData[(sy * imgW + sx) * 4 + 3] > 128) return true;
}
}
return false;
});
const totalPoints = gridPoints.length;
const allMasks = [];
// Helper: yield to browser so page stays responsive
const yieldToBrowser = () => new Promise(r => setTimeout(r, 0));
// Run mask decoder for each grid point
for (let i = 0; i < totalPoints; i++) {
if (onProgress) {
const pct = 30 + (i / totalPoints) * 50; // 30-80%
onProgress({
message: `Finding parts... ${i + 1}/${totalPoints}`,
progress: pct,
});
}
// Yield every few iterations to keep the page responsive
if (i % 3 === 0) await yieldToBrowser();
const [px, py] = gridPoints[i];
try {
const inputs = await processor(rawImage, {
input_points: [[[px, py]]],
input_labels: [[1]],
});
const outputs = await model(inputs);
const masks = await processor.post_process_masks(
outputs.pred_masks,
inputs.original_sizes,
inputs.reshaped_input_sizes,
);
// Get IoU scores - shape [1, numMasks]
const iouScores = outputs.iou_scores.data;
// Find best mask candidate
let bestIdx = 0;
let bestScore = iouScores[0];
for (let j = 1; j < iouScores.length; j++) {
if (iouScores[j] > bestScore) {
bestScore = iouScores[j];
bestIdx = j;
}
}
if (bestScore >= MIN_IOU_SCORE) {
// Extract the mask data for the best candidate
const maskTensor = masks[0][0]; // [numMasks, H, W]
const maskH = maskTensor.dims[1];
const maskW = maskTensor.dims[2];
const maskData = maskTensor.data;
const maskSize = maskH * maskW;
const offset = bestIdx * maskSize;
// Copy just this mask's data
const singleMask = new Float32Array(maskSize);
for (let k = 0; k < maskSize; k++) {
singleMask[k] = maskData[offset + k];
}
allMasks.push({
mask: singleMask,
maskW,
maskH,
score: bestScore,
});
}
// Dispose tensors
if (outputs.pred_masks.dispose) outputs.pred_masks.dispose();
if (outputs.iou_scores.dispose) outputs.iou_scores.dispose();
} catch (e) {
console.warn(`Grid point ${i} failed:`, e);
}
}
if (onProgress) onProgress({ message: 'Filtering results...', progress: 82 });
// Compute bounding box and area for each mask
const minArea = imgW * imgH * MIN_AREA_FRAC;
let candidates = allMasks.map((m, i) => {
const { bbox, area } = computeMaskStats(m.mask, m.maskW, m.maskH);
return { ...m, bbox, area, id: `seg-${i}` };
});
// Filter by area
candidates = candidates.filter(m => m.area >= minArea);
// NMS
candidates = nonMaxSuppression(candidates, NMS_IOU_THRESHOLD);
if (onProgress) onProgress({ message: 'Extracting parts...', progress: 88 });
// Crop each mask from the original image
const results = [];
for (let i = 0; i < candidates.length; i++) {
const c = candidates[i];
c.id = `seg-${i}`;
try {
c.croppedBlob = await extractMaskRegion(imageBlob, c);
results.push(c);
} catch (e) {
console.warn(`Failed to extract segment ${i}:`, e);
}
}
// Sort by area descending (largest first)
results.sort((a, b) => b.area - a.area);
// Dispose model
try { if (model.dispose) model.dispose(); } catch (_) {}
if (onProgress) onProgress({ message: 'Done!', progress: 100 });
return results;
}
// ---- Grid generation ----
function generateGrid(w, h, n) {
const points = [];
const stepX = w / (n + 1);
const stepY = h / (n + 1);
for (let row = 1; row <= n; row++) {
for (let col = 1; col <= n; col++) {
points.push([Math.round(col * stepX), Math.round(row * stepY)]);
}
}
return points;
}
// ---- Mask stats ----
function computeMaskStats(mask, w, h) {
let minX = w, minY = h, maxX = 0, maxY = 0;
let area = 0;
for (let y = 0; y < h; y++) {
for (let x = 0; x < w; x++) {
if (mask[y * w + x] > 0) {
area++;
if (x < minX) minX = x;
if (x > maxX) maxX = x;
if (y < minY) minY = y;
if (y > maxY) maxY = y;
}
}
}
return {
bbox: {
x: minX / w,
y: minY / h,
w: (maxX - minX + 1) / w,
h: (maxY - minY + 1) / h,
},
area,
};
}
// ---- Non-maximum suppression ----
function maskIoU(a, b) {
// Both masks must have same dimensions
const len = a.mask.length;
let intersection = 0, union = 0;
for (let i = 0; i < len; i++) {
const av = a.mask[i] > 0 ? 1 : 0;
const bv = b.mask[i] > 0 ? 1 : 0;
if (av && bv) intersection++;
if (av || bv) union++;
}
return union === 0 ? 0 : intersection / union;
}
function nonMaxSuppression(masks, threshold) {
// Sort by score descending
const sorted = [...masks].sort((a, b) => b.score - a.score);
const kept = [];
for (const candidate of sorted) {
let dominated = false;
for (const existing of kept) {
if (maskIoU(candidate, existing) > threshold) {
dominated = true;
break;
}
}
if (!dominated) {
kept.push(candidate);
}
}
return kept;
}
// ---- Extract mask region as cropped blob ----
async function extractMaskRegion(imageBlob, segment) {
const bitmap = await createImageBitmap(imageBlob);
const fullW = bitmap.width;
const fullH = bitmap.height;
// Convert normalized bbox to pixel coords with padding
const pad = 4;
const bx = Math.max(0, Math.floor(segment.bbox.x * fullW) - pad);
const by = Math.max(0, Math.floor(segment.bbox.y * fullH) - pad);
const bw = Math.min(fullW - bx, Math.ceil(segment.bbox.w * fullW) + pad * 2);
const bh = Math.min(fullH - by, Math.ceil(segment.bbox.h * fullH) + pad * 2);
const canvas = new OffscreenCanvas(bw, bh);
const ctx = canvas.getContext('2d');
ctx.drawImage(bitmap, bx, by, bw, bh, 0, 0, bw, bh);
bitmap.close();
// Apply mask as alpha
const imgData = ctx.getImageData(0, 0, bw, bh);
const scaleX = segment.maskW / fullW;
const scaleY = segment.maskH / fullH;
for (let y = 0; y < bh; y++) {
for (let x = 0; x < bw; x++) {
const mx = Math.min(Math.floor((bx + x) * scaleX), segment.maskW - 1);
const my = Math.min(Math.floor((by + y) * scaleY), segment.maskH - 1);
const maskVal = segment.mask[my * segment.maskW + mx] > 0 ? 1 : 0;
const idx = (y * bw + x) * 4;
// Multiply existing alpha with mask
imgData.data[idx + 3] = Math.round(imgData.data[idx + 3] * maskVal);
}
}
ctx.putImageData(imgData, 0, 0);
return canvas.convertToBlob({ type: 'image/png' });
}
// ---- Downscale ----
async function downscale(imageBlob, maxDim) {
const probe = await createImageBitmap(imageBlob);
const { width, height } = probe;
if (width <= maxDim && height <= maxDim) {
probe.close();
return imageBlob;
}
const ratio = Math.min(maxDim / width, maxDim / height);
const newW = Math.round(width * ratio);
const newH = Math.round(height * ratio);
const resized = await createImageBitmap(imageBlob, {
resizeWidth: newW, resizeHeight: newH, resizeQuality: 'medium',
});
probe.close();
const canvas = new OffscreenCanvas(newW, newH);
canvas.getContext('2d').drawImage(resized, 0, 0);
resized.close();
return canvas.convertToBlob({ type: 'image/png' });
}
// ---- Mask to polygon (for adjustment) ----
/**
* Convert a binary mask to a simplified polygon (normalized 0-1 coords).
* Uses border tracing and Douglas-Peucker simplification.
*/
export function maskToPolygon(mask, maskW, maskH, maxPoints = 30) {
// Find contour points using simple border following
const contour = traceContour(mask, maskW, maskH);
if (contour.length < 3) return contour;
// Simplify with Douglas-Peucker
const tolerance = Math.max(maskW, maskH) * 0.015;
let simplified = douglasPeucker(contour, tolerance);
// Cap points
while (simplified.length > maxPoints) {
simplified = douglasPeucker(simplified, tolerance * 1.5);
}
// Normalize to 0-1
return simplified.map(p => ({ x: p.x / maskW, y: p.y / maskH }));
}
function traceContour(mask, w, h) {
// Find first border pixel
let startX = -1, startY = -1;
outer: for (let y = 0; y < h; y++) {
for (let x = 0; x < w; x++) {
if (mask[y * w + x] > 0) {
startX = x;
startY = y;
break outer;
}
}
}
if (startX < 0) return [];
const contour = [];
const dirs = [
[1, 0], [1, 1], [0, 1], [-1, 1],
[-1, 0], [-1, -1], [0, -1], [1, -1],
];
let cx = startX, cy = startY;
let dir = 0;
const maxSteps = w * h;
for (let step = 0; step < maxSteps; step++) {
contour.push({ x: cx, y: cy });
// Look for next border pixel
let found = false;
const startDir = (dir + 5) % 8; // turn back to find outline
for (let i = 0; i < 8; i++) {
const d = (startDir + i) % 8;
const nx = cx + dirs[d][0];
const ny = cy + dirs[d][1];
if (nx >= 0 && nx < w && ny >= 0 && ny < h && mask[ny * w + nx] > 0) {
cx = nx;
cy = ny;
dir = d;
found = true;
break;
}
}
if (!found || (cx === startX && cy === startY && step > 2)) break;
}
// Subsample to avoid too many points
if (contour.length > 200) {
const step = Math.ceil(contour.length / 200);
const subsampled = [];
for (let i = 0; i < contour.length; i += step) {
subsampled.push(contour[i]);
}
return subsampled;
}
return contour;
}
function douglasPeucker(points, epsilon) {
if (points.length <= 2) return points;
let maxDist = 0;
let maxIdx = 0;
const first = points[0];
const last = points[points.length - 1];
for (let i = 1; i < points.length - 1; i++) {
const d = pointToLineDist(points[i], first, last);
if (d > maxDist) {
maxDist = d;
maxIdx = i;
}
}
if (maxDist > epsilon) {
const left = douglasPeucker(points.slice(0, maxIdx + 1), epsilon);
const right = douglasPeucker(points.slice(maxIdx), epsilon);
return [...left.slice(0, -1), ...right];
}
return [first, last];
}
function pointToLineDist(p, a, b) {
const dx = b.x - a.x;
const dy = b.y - a.y;
const len2 = dx * dx + dy * dy;
if (len2 === 0) return Math.sqrt((p.x - a.x) ** 2 + (p.y - a.y) ** 2);
const t = Math.max(0, Math.min(1, ((p.x - a.x) * dx + (p.y - a.y) * dy) / len2));
const projX = a.x + t * dx;
const projY = a.y + t * dy;
return Math.sqrt((p.x - projX) ** 2 + (p.y - projY) ** 2);
}
|