@camstack/addon-pipeline 1.1.53 → 1.1.55
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/audio-analyzer/index.js +1 -1
- package/dist/audio-analyzer/index.mjs +1 -1
- package/dist/detection-pipeline/index.js +289 -1469
- package/dist/detection-pipeline/index.mjs +269 -1449
- package/dist/{dist-DWb62H5U.js → dist-DI57FC8K.js} +167 -8
- package/dist/{dist-BalxNwOt.mjs → dist-RWGGPwVx.mjs} +167 -8
- package/dist/motion-wasm/index.js +1 -1
- package/dist/motion-wasm/index.mjs +1 -1
- package/dist/pipeline-runner/index.js +772 -21
- package/dist/pipeline-runner/index.mjs +772 -22
- package/dist/recorder/index.js +1 -1
- package/dist/recorder/index.mjs +1 -1
- package/dist/{remote-source-plane-CZpzIVro.js → remote-source-plane-CHgvzzA6.js} +1 -1
- package/dist/{remote-source-plane-BCJW5CvF.mjs → remote-source-plane-DU0aRSPv.mjs} +1 -1
- package/dist/session-decode/decode-worker-child.js +538 -35
- package/dist/session-decode/decode-worker-child.mjs +538 -35
- package/dist/step-definitions-CNBFKjZe.js +1514 -0
- package/dist/step-definitions-CP9kVSml.mjs +1479 -0
- package/dist/stream-broker/_stub.js +2 -2
- package/dist/stream-broker/{_virtual_mf-localSharedImportMap___mfe_internal__addon_stream_broker_widgets-DIN3CcRP.mjs → _virtual_mf-localSharedImportMap___mfe_internal__addon_stream_broker_widgets-mDCPzmT3.mjs} +3 -3
- package/dist/stream-broker/_virtual_mf___mfe_internal__addon_stream_broker_widgets__loadShare___mf_0_camstack_mf_1_types__loadShare__.js-B3gTdHEh.mjs +26 -0
- package/dist/stream-broker/{_virtual_mf___mfe_internal__addon_stream_broker_widgets__loadShare___mf_0_camstack_mf_1_ui_mf_2_library__loadShare__.js-C9fwKMfg.mjs → _virtual_mf___mfe_internal__addon_stream_broker_widgets__loadShare___mf_0_camstack_mf_1_ui_mf_2_library__loadShare__.js-Dn_pxzP-.mjs} +1 -1
- package/dist/stream-broker/{hostInit-CGUEq--X.mjs → hostInit-GM_CI22k.mjs} +3 -3
- package/dist/stream-broker/index.js +2 -2
- package/dist/stream-broker/index.mjs +2 -2
- package/dist/stream-broker/remoteEntry.js +1 -1
- package/dist/{worker-protocol-pk7qdYXt.mjs → worker-protocol-CyVJTZEO.mjs} +7 -0
- package/dist/{worker-protocol-BCfO8gUF.js → worker-protocol-PP4jKHHJ.js} +7 -0
- package/embed-dist/assets/{MaskShapeCanvas-DI4BY7W2-CqOe8eYa.js → MaskShapeCanvas-DI4BY7W2-BChW0ntM.js} +1 -1
- package/embed-dist/assets/{MotionZonesSettings-NcxxQN8r-CZyLeUnd.js → MotionZonesSettings-NcxxQN8r-DDzqEbSe.js} +1 -1
- package/embed-dist/assets/{PrivacyMaskSettings-APgPLF7p-Cn0hGZnu.js → PrivacyMaskSettings-APgPLF7p-B879GXaf.js} +1 -1
- package/embed-dist/assets/{index-DRl4XYjA.js → index-CNjQ5rAE.js} +10 -10
- package/embed-dist/index.html +1 -1
- package/package.json +1 -1
- package/dist/stream-broker/_virtual_mf___mfe_internal__addon_stream_broker_widgets__loadShare___mf_0_camstack_mf_1_types__loadShare__.js-o4tu_xuc.mjs +0 -26
|
@@ -1,5 +1,6 @@
|
|
|
1
1
|
import { t as __require } from "../chunk-BdkLduGY.mjs";
|
|
2
|
-
import { H as createEvent, I as errMsg, L as BaseAddon, O as pipelineExecutorCapability, P as supportedRuntimes$1,
|
|
2
|
+
import { D as nodePin, H as createEvent, I as errMsg, L as BaseAddon, O as pipelineExecutorCapability, P as supportedRuntimes$1, U as hydrateSchema, V as EventCategory, X as sleep, b as detectionPipelineCapability, f as YAMNET_TO_MACRO, j as runtimeDevices$1, q as parseJsonUnknown, t as APPLE_SA_TO_MACRO, v as defaultDeviceFor$1, x as evaluateZoneRules } from "../dist-RWGGPwVx.mjs";
|
|
3
|
+
import { a as getStepDefinition, i as getStep, n as ALL_STEPS, o as resolveModelForFormat, r as getDefaultModelForFormat, t as ALL_PIPELINE_STEPS } from "../step-definitions-CP9kVSml.mjs";
|
|
3
4
|
import { t as pickNodePlatformArch } from "../node-topology-platform-BkR_k6WT.mjs";
|
|
4
5
|
import { a as isModelDownloaded, i as ensureModel, n as deleteModelFromDisk } from "../model-download-service-C-IHWnXx-BPy6aoAx.mjs";
|
|
5
6
|
import { FrameRingReaderCache } from "@camstack/shm-ring";
|
|
@@ -290,1437 +291,6 @@ function projectNodeEngine(store, nodeId) {
|
|
|
290
291
|
return out;
|
|
291
292
|
}
|
|
292
293
|
//#endregion
|
|
293
|
-
//#region src/detection-pipeline/registry/model-catalogs.ts
|
|
294
|
-
var HF_REPO = "camstack/camstack-models";
|
|
295
|
-
var HF_SCRYPTED = "scrypted/plugin-models";
|
|
296
|
-
var hf = (path) => hfModelUrl(HF_REPO, path);
|
|
297
|
-
var hfScrypted = (path) => hfModelUrl(HF_SCRYPTED, path);
|
|
298
|
-
/**
|
|
299
|
-
* Build an OpenVINO format entry (always python runtime).
|
|
300
|
-
*
|
|
301
|
-
* OpenVINO IR is a two-file bundle: a `.xml` topology + a sibling `.bin`
|
|
302
|
-
* weights file with the same basename. We declare the `.bin` in `files` so
|
|
303
|
-
* the (format-agnostic) downloader fetches it alongside the `.xml` — without
|
|
304
|
-
* the weights, OpenVINO compile fails with "Empty weights data in bin file".
|
|
305
|
-
* A plain `.onnx` run through the OpenVINO runtime (e.g. yamnet) has no
|
|
306
|
-
* sibling, so none is added.
|
|
307
|
-
*/
|
|
308
|
-
var ovFormat = (url, sizeMB) => {
|
|
309
|
-
const base = url.split("/").pop() ?? "";
|
|
310
|
-
const files = base.endsWith(".xml") ? [base.replace(/\.xml$/, ".bin")] : void 0;
|
|
311
|
-
return {
|
|
312
|
-
url,
|
|
313
|
-
sizeMB,
|
|
314
|
-
runtimes: ["python"],
|
|
315
|
-
...files ? { files } : {}
|
|
316
|
-
};
|
|
317
|
-
};
|
|
318
|
-
/**
|
|
319
|
-
* Build a precision-variant catalog entry (OpenVINO-only) derived from a base
|
|
320
|
-
* detection model. fp16 halves the weights (Intel iGPU/NPU sweet spot); int8 is
|
|
321
|
-
* NNCF post-training-quantized (~4× smaller, fastest on CPU/iGPU at a small
|
|
322
|
-
* accuracy cost). The IRs live next to the base `.xml` on HF as
|
|
323
|
-
* `camstack-<id>-<precision>.xml`. Lets an operator scale the model to the node
|
|
324
|
-
* (e.g. yolo26x-int8 on a 265K, yolo26n-int8 on an N100).
|
|
325
|
-
*/
|
|
326
|
-
var ovPrecisionVariant = (baseId, ovDir, baseName, precision, sizeMB, legacy = false) => ({
|
|
327
|
-
id: `${baseId}-${precision}`,
|
|
328
|
-
name: `${baseName} (${precision.toUpperCase()})`,
|
|
329
|
-
description: `${baseName} — OpenVINO ${precision.toUpperCase()} variant for Intel iGPU/NPU; scale by hardware`,
|
|
330
|
-
inputSize: {
|
|
331
|
-
width: 640,
|
|
332
|
-
height: 640
|
|
333
|
-
},
|
|
334
|
-
labels: [],
|
|
335
|
-
preprocessMode: "letterbox",
|
|
336
|
-
formats: { openvino: ovFormat(hf(`${ovDir}/camstack-${baseId}-${precision}.xml`), sizeMB) },
|
|
337
|
-
...legacy ? { legacy: true } : {},
|
|
338
|
-
...precision === "int8" ? { group: {
|
|
339
|
-
family: baseId.slice(0, -1),
|
|
340
|
-
tier: baseId.slice(-1),
|
|
341
|
-
precision: "int8"
|
|
342
|
-
} } : {}
|
|
343
|
-
});
|
|
344
|
-
var MLPACKAGE_FILES = [
|
|
345
|
-
"Manifest.json",
|
|
346
|
-
"Data/com.apple.CoreML/model.mlmodel",
|
|
347
|
-
"Data/com.apple.CoreML/weights/weight.bin"
|
|
348
|
-
];
|
|
349
|
-
var YOLO26_TIER_NAME = {
|
|
350
|
-
n: "Nano",
|
|
351
|
-
s: "Small",
|
|
352
|
-
m: "Medium",
|
|
353
|
-
l: "Large"
|
|
354
|
-
};
|
|
355
|
-
/**
|
|
356
|
-
* Reduced-input YOLO26 variants (2026-07-11). A lower square input (320 / 256)
|
|
357
|
-
* is the biggest, cheapest latency lever measured across every runtime — Mac
|
|
358
|
-
* ANE, Intel Ultra NPU, and especially the N100 — at a small-object accuracy
|
|
359
|
-
* cost. Each resolution ships a base (fp32) entry (onnx + coreml + openvino)
|
|
360
|
-
* grouped under `{family, tier, resolution}`, plus an OpenVINO INT8 entry (best
|
|
361
|
-
* on Intel/N100) grouped with `precision:'int8'`. The grouped selector surfaces
|
|
362
|
-
* `resolution` as a variant axis alongside precision.
|
|
363
|
-
*/
|
|
364
|
-
var yolo26Reduced = (tier, res, sizes) => {
|
|
365
|
-
const name = YOLO26_TIER_NAME[tier] ?? tier.toUpperCase();
|
|
366
|
-
return [{
|
|
367
|
-
id: `yolo26${tier}-${res}`,
|
|
368
|
-
group: {
|
|
369
|
-
family: "yolo26",
|
|
370
|
-
tier,
|
|
371
|
-
resolution: res
|
|
372
|
-
},
|
|
373
|
-
name: `YOLO26 ${name} @${res}`,
|
|
374
|
-
description: `YOLO26 ${name}, reduced ${res}×${res} input — faster on ANE / hub NPU / N100 at a small-object accuracy cost`,
|
|
375
|
-
inputSize: {
|
|
376
|
-
width: res,
|
|
377
|
-
height: res
|
|
378
|
-
},
|
|
379
|
-
labels: [],
|
|
380
|
-
preprocessMode: "letterbox",
|
|
381
|
-
formats: {
|
|
382
|
-
onnx: {
|
|
383
|
-
url: hf(`objectDetection/yolo26/onnx/camstack-yolo26${tier}-${res}.onnx`),
|
|
384
|
-
sizeMB: sizes.onnx
|
|
385
|
-
},
|
|
386
|
-
coreml: {
|
|
387
|
-
url: hf(`objectDetection/yolo26/coreml/camstack-yolo26${tier}-${res}.mlpackage`),
|
|
388
|
-
sizeMB: sizes.coreml,
|
|
389
|
-
isDirectory: true,
|
|
390
|
-
files: [...MLPACKAGE_FILES],
|
|
391
|
-
runtimes: ["python"]
|
|
392
|
-
},
|
|
393
|
-
openvino: ovFormat(hf(`objectDetection/yolo26/openvino/camstack-yolo26${tier}-${res}.xml`), sizes.ovFp32)
|
|
394
|
-
}
|
|
395
|
-
}, {
|
|
396
|
-
id: `yolo26${tier}-${res}-int8`,
|
|
397
|
-
group: {
|
|
398
|
-
family: "yolo26",
|
|
399
|
-
tier,
|
|
400
|
-
precision: "int8",
|
|
401
|
-
resolution: res
|
|
402
|
-
},
|
|
403
|
-
name: `YOLO26 ${name} @${res} (INT8)`,
|
|
404
|
-
description: `YOLO26 ${name} @${res} — OpenVINO INT8 for Intel iGPU/CPU; the N100's fastest config`,
|
|
405
|
-
inputSize: {
|
|
406
|
-
width: res,
|
|
407
|
-
height: res
|
|
408
|
-
},
|
|
409
|
-
labels: [],
|
|
410
|
-
preprocessMode: "letterbox",
|
|
411
|
-
formats: { openvino: ovFormat(hf(`objectDetection/yolo26/openvino/camstack-yolo26${tier}-${res}-int8.xml`), sizes.ovInt8) }
|
|
412
|
-
}];
|
|
413
|
-
};
|
|
414
|
-
var OBJECT_DETECTION_MODELS = [
|
|
415
|
-
{
|
|
416
|
-
id: "yolov9t",
|
|
417
|
-
legacy: true,
|
|
418
|
-
name: "YOLOv9 Tiny",
|
|
419
|
-
description: "YOLOv9 Tiny — ultra-lightweight next-gen detector",
|
|
420
|
-
inputSize: {
|
|
421
|
-
width: 640,
|
|
422
|
-
height: 640
|
|
423
|
-
},
|
|
424
|
-
labels: [],
|
|
425
|
-
preprocessMode: "letterbox",
|
|
426
|
-
formats: {
|
|
427
|
-
onnx: {
|
|
428
|
-
url: hf("objectDetection/yolov9/onnx/camstack-yolov9t.onnx"),
|
|
429
|
-
sizeMB: 8
|
|
430
|
-
},
|
|
431
|
-
coreml: {
|
|
432
|
-
url: hf("objectDetection/yolov9/coreml/camstack-yolov9t.mlpackage"),
|
|
433
|
-
sizeMB: 4,
|
|
434
|
-
isDirectory: true,
|
|
435
|
-
files: [...MLPACKAGE_FILES],
|
|
436
|
-
runtimes: ["python"]
|
|
437
|
-
},
|
|
438
|
-
openvino: ovFormat(hf("objectDetection/yolov9/openvino/camstack-yolov9t.xml"), 6)
|
|
439
|
-
}
|
|
440
|
-
},
|
|
441
|
-
{
|
|
442
|
-
id: "yolov9s",
|
|
443
|
-
legacy: true,
|
|
444
|
-
name: "YOLOv9 Small",
|
|
445
|
-
description: "YOLOv9 Small — improved efficiency over YOLOv8s",
|
|
446
|
-
inputSize: {
|
|
447
|
-
width: 640,
|
|
448
|
-
height: 640
|
|
449
|
-
},
|
|
450
|
-
labels: [],
|
|
451
|
-
preprocessMode: "letterbox",
|
|
452
|
-
formats: {
|
|
453
|
-
onnx: {
|
|
454
|
-
url: hf("objectDetection/yolov9/onnx/camstack-yolov9s.onnx"),
|
|
455
|
-
sizeMB: 28
|
|
456
|
-
},
|
|
457
|
-
coreml: {
|
|
458
|
-
url: hf("objectDetection/yolov9/coreml/camstack-yolov9s.mlpackage"),
|
|
459
|
-
sizeMB: 14,
|
|
460
|
-
isDirectory: true,
|
|
461
|
-
files: [...MLPACKAGE_FILES],
|
|
462
|
-
runtimes: ["python"]
|
|
463
|
-
},
|
|
464
|
-
openvino: ovFormat(hf("objectDetection/yolov9/openvino/camstack-yolov9s.xml"), 16)
|
|
465
|
-
}
|
|
466
|
-
},
|
|
467
|
-
{
|
|
468
|
-
id: "yolov9c",
|
|
469
|
-
legacy: true,
|
|
470
|
-
name: "YOLOv9 C",
|
|
471
|
-
description: "YOLOv9 C — high-accuracy compact model",
|
|
472
|
-
inputSize: {
|
|
473
|
-
width: 640,
|
|
474
|
-
height: 640
|
|
475
|
-
},
|
|
476
|
-
labels: [],
|
|
477
|
-
preprocessMode: "letterbox",
|
|
478
|
-
formats: {
|
|
479
|
-
onnx: {
|
|
480
|
-
url: hf("objectDetection/yolov9/onnx/camstack-yolov9c.onnx"),
|
|
481
|
-
sizeMB: 97
|
|
482
|
-
},
|
|
483
|
-
coreml: {
|
|
484
|
-
url: hf("objectDetection/yolov9/coreml/camstack-yolov9c.mlpackage"),
|
|
485
|
-
sizeMB: 48,
|
|
486
|
-
isDirectory: true,
|
|
487
|
-
files: [...MLPACKAGE_FILES],
|
|
488
|
-
runtimes: ["python"]
|
|
489
|
-
},
|
|
490
|
-
openvino: ovFormat(hf("objectDetection/yolov9/openvino/camstack-yolov9c.xml"), 49)
|
|
491
|
-
}
|
|
492
|
-
},
|
|
493
|
-
{
|
|
494
|
-
id: "yolo26n",
|
|
495
|
-
group: {
|
|
496
|
-
family: "yolo26",
|
|
497
|
-
tier: "n"
|
|
498
|
-
},
|
|
499
|
-
name: "YOLO26 Nano",
|
|
500
|
-
description: "YOLO26 Nano — latest generation ultra-lightweight detector with attention mechanism",
|
|
501
|
-
inputSize: {
|
|
502
|
-
width: 640,
|
|
503
|
-
height: 640
|
|
504
|
-
},
|
|
505
|
-
labels: [],
|
|
506
|
-
preprocessMode: "letterbox",
|
|
507
|
-
formats: {
|
|
508
|
-
onnx: {
|
|
509
|
-
url: hf("objectDetection/yolo26/onnx/camstack-yolo26n.onnx"),
|
|
510
|
-
sizeMB: 10
|
|
511
|
-
},
|
|
512
|
-
coreml: {
|
|
513
|
-
url: hf("objectDetection/yolo26/coreml/camstack-yolo26n.mlpackage"),
|
|
514
|
-
sizeMB: 5,
|
|
515
|
-
isDirectory: true,
|
|
516
|
-
files: [...MLPACKAGE_FILES],
|
|
517
|
-
runtimes: ["python"]
|
|
518
|
-
},
|
|
519
|
-
openvino: ovFormat(hf("objectDetection/yolo26/openvino/camstack-yolo26n.xml"), 9)
|
|
520
|
-
}
|
|
521
|
-
},
|
|
522
|
-
{
|
|
523
|
-
id: "yolo26s",
|
|
524
|
-
group: {
|
|
525
|
-
family: "yolo26",
|
|
526
|
-
tier: "s"
|
|
527
|
-
},
|
|
528
|
-
name: "YOLO26 Small",
|
|
529
|
-
description: "YOLO26 Small — balanced speed/accuracy with attention-based architecture",
|
|
530
|
-
inputSize: {
|
|
531
|
-
width: 640,
|
|
532
|
-
height: 640
|
|
533
|
-
},
|
|
534
|
-
labels: [],
|
|
535
|
-
preprocessMode: "letterbox",
|
|
536
|
-
formats: {
|
|
537
|
-
onnx: {
|
|
538
|
-
url: hf("objectDetection/yolo26/onnx/camstack-yolo26s.onnx"),
|
|
539
|
-
sizeMB: 37
|
|
540
|
-
},
|
|
541
|
-
coreml: {
|
|
542
|
-
url: hf("objectDetection/yolo26/coreml/camstack-yolo26s.mlpackage"),
|
|
543
|
-
sizeMB: 18,
|
|
544
|
-
isDirectory: true,
|
|
545
|
-
files: [...MLPACKAGE_FILES],
|
|
546
|
-
runtimes: ["python"]
|
|
547
|
-
},
|
|
548
|
-
openvino: ovFormat(hf("objectDetection/yolo26/openvino/camstack-yolo26s.xml"), 36)
|
|
549
|
-
}
|
|
550
|
-
},
|
|
551
|
-
{
|
|
552
|
-
id: "yolo26m",
|
|
553
|
-
group: {
|
|
554
|
-
family: "yolo26",
|
|
555
|
-
tier: "m"
|
|
556
|
-
},
|
|
557
|
-
name: "YOLO26 Medium",
|
|
558
|
-
description: "YOLO26 Medium — high accuracy general-purpose detector",
|
|
559
|
-
inputSize: {
|
|
560
|
-
width: 640,
|
|
561
|
-
height: 640
|
|
562
|
-
},
|
|
563
|
-
labels: [],
|
|
564
|
-
preprocessMode: "letterbox",
|
|
565
|
-
formats: {
|
|
566
|
-
onnx: {
|
|
567
|
-
url: hf("objectDetection/yolo26/onnx/camstack-yolo26m.onnx"),
|
|
568
|
-
sizeMB: 78
|
|
569
|
-
},
|
|
570
|
-
coreml: {
|
|
571
|
-
url: hf("objectDetection/yolo26/coreml/camstack-yolo26m.mlpackage"),
|
|
572
|
-
sizeMB: 39,
|
|
573
|
-
isDirectory: true,
|
|
574
|
-
files: [...MLPACKAGE_FILES],
|
|
575
|
-
runtimes: ["python"]
|
|
576
|
-
},
|
|
577
|
-
openvino: ovFormat(hf("objectDetection/yolo26/openvino/camstack-yolo26m.xml"), 78)
|
|
578
|
-
}
|
|
579
|
-
},
|
|
580
|
-
{
|
|
581
|
-
id: "yolo26l",
|
|
582
|
-
group: {
|
|
583
|
-
family: "yolo26",
|
|
584
|
-
tier: "l"
|
|
585
|
-
},
|
|
586
|
-
name: "YOLO26 Large",
|
|
587
|
-
description: "YOLO26 Large — maximum accuracy for demanding deployments",
|
|
588
|
-
inputSize: {
|
|
589
|
-
width: 640,
|
|
590
|
-
height: 640
|
|
591
|
-
},
|
|
592
|
-
labels: [],
|
|
593
|
-
preprocessMode: "letterbox",
|
|
594
|
-
formats: {
|
|
595
|
-
onnx: {
|
|
596
|
-
url: hf("objectDetection/yolo26/onnx/camstack-yolo26l.onnx"),
|
|
597
|
-
sizeMB: 95
|
|
598
|
-
},
|
|
599
|
-
coreml: {
|
|
600
|
-
url: hf("objectDetection/yolo26/coreml/camstack-yolo26l.mlpackage"),
|
|
601
|
-
sizeMB: 48,
|
|
602
|
-
isDirectory: true,
|
|
603
|
-
files: [...MLPACKAGE_FILES],
|
|
604
|
-
runtimes: ["python"]
|
|
605
|
-
},
|
|
606
|
-
openvino: ovFormat(hf("objectDetection/yolo26/openvino/camstack-yolo26l.xml"), 95)
|
|
607
|
-
}
|
|
608
|
-
},
|
|
609
|
-
{
|
|
610
|
-
id: "yolo26x",
|
|
611
|
-
legacy: true,
|
|
612
|
-
name: "YOLO26 XLarge",
|
|
613
|
-
description: "YOLO26 XLarge — highest accuracy, attention-based architecture",
|
|
614
|
-
inputSize: {
|
|
615
|
-
width: 640,
|
|
616
|
-
height: 640
|
|
617
|
-
},
|
|
618
|
-
labels: [],
|
|
619
|
-
preprocessMode: "letterbox",
|
|
620
|
-
formats: {
|
|
621
|
-
onnx: {
|
|
622
|
-
url: hf("objectDetection/yolo26/onnx/camstack-yolo26x.onnx"),
|
|
623
|
-
sizeMB: 213
|
|
624
|
-
},
|
|
625
|
-
coreml: {
|
|
626
|
-
url: hf("objectDetection/yolo26/coreml/camstack-yolo26x.mlpackage"),
|
|
627
|
-
sizeMB: 107,
|
|
628
|
-
isDirectory: true,
|
|
629
|
-
files: [...MLPACKAGE_FILES],
|
|
630
|
-
runtimes: ["python"]
|
|
631
|
-
},
|
|
632
|
-
openvino: ovFormat(hf("objectDetection/yolo26/openvino/camstack-yolo26x.xml"), 213)
|
|
633
|
-
}
|
|
634
|
-
},
|
|
635
|
-
{
|
|
636
|
-
id: "scrypted-yolov9t-relu",
|
|
637
|
-
legacy: true,
|
|
638
|
-
name: "YOLOv9 Tiny ReLU",
|
|
639
|
-
description: "YOLOv9t ReLU — optimized for int8 quantization",
|
|
640
|
-
inputSize: {
|
|
641
|
-
width: 320,
|
|
642
|
-
height: 320
|
|
643
|
-
},
|
|
644
|
-
labels: [],
|
|
645
|
-
preprocessMode: "letterbox",
|
|
646
|
-
formats: {
|
|
647
|
-
onnx: {
|
|
648
|
-
url: hfScrypted("onnx/scrypted_yolov9t_relu/scrypted_yolov9t_relu.onnx"),
|
|
649
|
-
sizeMB: 8.1
|
|
650
|
-
},
|
|
651
|
-
coreml: {
|
|
652
|
-
url: hfScrypted("coreml/scrypted_yolov9t_relu/scrypted_yolov9t_relu.mlpackage"),
|
|
653
|
-
sizeMB: 4,
|
|
654
|
-
isDirectory: true,
|
|
655
|
-
files: [...MLPACKAGE_FILES],
|
|
656
|
-
runtimes: ["python"]
|
|
657
|
-
},
|
|
658
|
-
openvino: ovFormat(hf("objectDetection/scrypted-yolov9-relu/openvino/scrypted_yolov9t_relu.xml"), 6)
|
|
659
|
-
}
|
|
660
|
-
},
|
|
661
|
-
{
|
|
662
|
-
id: "scrypted-yolov9s-relu",
|
|
663
|
-
legacy: true,
|
|
664
|
-
name: "YOLOv9 Small ReLU",
|
|
665
|
-
description: "YOLOv9s ReLU — improved efficiency, int8 ready",
|
|
666
|
-
inputSize: {
|
|
667
|
-
width: 320,
|
|
668
|
-
height: 320
|
|
669
|
-
},
|
|
670
|
-
labels: [],
|
|
671
|
-
preprocessMode: "letterbox",
|
|
672
|
-
formats: {
|
|
673
|
-
onnx: {
|
|
674
|
-
url: hfScrypted("onnx/scrypted_yolov9s_relu/scrypted_yolov9s_relu.onnx"),
|
|
675
|
-
sizeMB: 27.9
|
|
676
|
-
},
|
|
677
|
-
coreml: {
|
|
678
|
-
url: hfScrypted("coreml/scrypted_yolov9s_relu/scrypted_yolov9s_relu.mlpackage"),
|
|
679
|
-
sizeMB: 14,
|
|
680
|
-
isDirectory: true,
|
|
681
|
-
files: [...MLPACKAGE_FILES],
|
|
682
|
-
runtimes: ["python"]
|
|
683
|
-
},
|
|
684
|
-
openvino: ovFormat(hf("objectDetection/scrypted-yolov9-relu/openvino/scrypted_yolov9s_relu.xml"), 16)
|
|
685
|
-
}
|
|
686
|
-
},
|
|
687
|
-
{
|
|
688
|
-
id: "scrypted-yolov9c-relu",
|
|
689
|
-
legacy: true,
|
|
690
|
-
name: "YOLOv9 C ReLU",
|
|
691
|
-
description: "YOLOv9c ReLU — high-accuracy, int8 ready",
|
|
692
|
-
inputSize: {
|
|
693
|
-
width: 320,
|
|
694
|
-
height: 320
|
|
695
|
-
},
|
|
696
|
-
labels: [],
|
|
697
|
-
preprocessMode: "letterbox",
|
|
698
|
-
formats: {
|
|
699
|
-
onnx: {
|
|
700
|
-
url: hfScrypted("onnx/scrypted_yolov9c_relu/scrypted_yolov9c_relu.onnx"),
|
|
701
|
-
sizeMB: 96.9
|
|
702
|
-
},
|
|
703
|
-
coreml: {
|
|
704
|
-
url: hfScrypted("coreml/scrypted_yolov9c_relu/scrypted_yolov9c_relu.mlpackage"),
|
|
705
|
-
sizeMB: 48,
|
|
706
|
-
isDirectory: true,
|
|
707
|
-
files: [...MLPACKAGE_FILES],
|
|
708
|
-
runtimes: ["python"]
|
|
709
|
-
},
|
|
710
|
-
openvino: ovFormat(hf("objectDetection/scrypted-yolov9-relu/openvino/scrypted_yolov9c_relu.xml"), 49)
|
|
711
|
-
}
|
|
712
|
-
},
|
|
713
|
-
{
|
|
714
|
-
id: "scrypted-yolov9m-relu",
|
|
715
|
-
legacy: true,
|
|
716
|
-
name: "YOLOv9 M ReLU",
|
|
717
|
-
description: "YOLOv9m ReLU — medium, int8 ready",
|
|
718
|
-
inputSize: {
|
|
719
|
-
width: 320,
|
|
720
|
-
height: 320
|
|
721
|
-
},
|
|
722
|
-
labels: [],
|
|
723
|
-
preprocessMode: "letterbox",
|
|
724
|
-
formats: {
|
|
725
|
-
onnx: {
|
|
726
|
-
url: hfScrypted("onnx/scrypted_yolov9m_relu/scrypted_yolov9m_relu.onnx"),
|
|
727
|
-
sizeMB: 76.6
|
|
728
|
-
},
|
|
729
|
-
coreml: {
|
|
730
|
-
url: hfScrypted("coreml/scrypted_yolov9m_relu/scrypted_yolov9m_relu.mlpackage"),
|
|
731
|
-
sizeMB: 38,
|
|
732
|
-
isDirectory: true,
|
|
733
|
-
files: [...MLPACKAGE_FILES],
|
|
734
|
-
runtimes: ["python"]
|
|
735
|
-
},
|
|
736
|
-
openvino: ovFormat(hf("objectDetection/scrypted-yolov9-relu/openvino/scrypted_yolov9m_relu.xml"), 38)
|
|
737
|
-
}
|
|
738
|
-
},
|
|
739
|
-
ovPrecisionVariant("yolov9t", "objectDetection/yolov9/openvino", "YOLOv9 Tiny", "fp16", 5, true),
|
|
740
|
-
ovPrecisionVariant("yolov9t", "objectDetection/yolov9/openvino", "YOLOv9 Tiny", "int8", 3, true),
|
|
741
|
-
ovPrecisionVariant("yolov9s", "objectDetection/yolov9/openvino", "YOLOv9 Small", "fp16", 15, true),
|
|
742
|
-
ovPrecisionVariant("yolov9s", "objectDetection/yolov9/openvino", "YOLOv9 Small", "int8", 8, true),
|
|
743
|
-
ovPrecisionVariant("yolo26n", "objectDetection/yolo26/openvino", "YOLO26 Nano", "fp16", 5, true),
|
|
744
|
-
ovPrecisionVariant("yolo26n", "objectDetection/yolo26/openvino", "YOLO26 Nano", "int8", 3),
|
|
745
|
-
ovPrecisionVariant("yolo26s", "objectDetection/yolo26/openvino", "YOLO26 Small", "fp16", 19, true),
|
|
746
|
-
ovPrecisionVariant("yolo26s", "objectDetection/yolo26/openvino", "YOLO26 Small", "int8", 10),
|
|
747
|
-
ovPrecisionVariant("yolo26m", "objectDetection/yolo26/openvino", "YOLO26 Medium", "fp16", 41, true),
|
|
748
|
-
ovPrecisionVariant("yolo26m", "objectDetection/yolo26/openvino", "YOLO26 Medium", "int8", 21),
|
|
749
|
-
ovPrecisionVariant("yolo26l", "objectDetection/yolo26/openvino", "YOLO26 Large", "fp16", 50, true),
|
|
750
|
-
ovPrecisionVariant("yolo26l", "objectDetection/yolo26/openvino", "YOLO26 Large", "int8", 25),
|
|
751
|
-
...yolo26Reduced("n", 320, {
|
|
752
|
-
onnx: 10.47,
|
|
753
|
-
coreml: 5.05,
|
|
754
|
-
ovFp32: 10.05,
|
|
755
|
-
ovInt8: 3.29
|
|
756
|
-
}),
|
|
757
|
-
...yolo26Reduced("n", 256, {
|
|
758
|
-
onnx: 10.47,
|
|
759
|
-
coreml: 5.05,
|
|
760
|
-
ovFp32: 10.05,
|
|
761
|
-
ovInt8: 3.29
|
|
762
|
-
}),
|
|
763
|
-
...yolo26Reduced("s", 320, {
|
|
764
|
-
onnx: 38.82,
|
|
765
|
-
coreml: 19.23,
|
|
766
|
-
ovFp32: 38.4,
|
|
767
|
-
ovInt8: 10.43
|
|
768
|
-
}),
|
|
769
|
-
...yolo26Reduced("s", 256, {
|
|
770
|
-
onnx: 38.82,
|
|
771
|
-
coreml: 19.23,
|
|
772
|
-
ovFp32: 38.4,
|
|
773
|
-
ovInt8: 10.43
|
|
774
|
-
}),
|
|
775
|
-
...yolo26Reduced("m", 320, {
|
|
776
|
-
onnx: 83.31,
|
|
777
|
-
coreml: 41.07,
|
|
778
|
-
ovFp32: 82.09,
|
|
779
|
-
ovInt8: 21.47
|
|
780
|
-
}),
|
|
781
|
-
...yolo26Reduced("m", 256, {
|
|
782
|
-
onnx: 83.31,
|
|
783
|
-
coreml: 41.07,
|
|
784
|
-
ovFp32: 82.09,
|
|
785
|
-
ovInt8: 21.47
|
|
786
|
-
}),
|
|
787
|
-
...yolo26Reduced("l", 320, {
|
|
788
|
-
onnx: 101.74,
|
|
789
|
-
coreml: 49.97,
|
|
790
|
-
ovFp32: 99.85,
|
|
791
|
-
ovInt8: 26.31
|
|
792
|
-
}),
|
|
793
|
-
...yolo26Reduced("l", 256, {
|
|
794
|
-
onnx: 101.74,
|
|
795
|
-
coreml: 49.97,
|
|
796
|
-
ovFp32: 99.85,
|
|
797
|
-
ovInt8: 26.31
|
|
798
|
-
}),
|
|
799
|
-
ovPrecisionVariant("yolo26x", "objectDetection/yolo26/openvino", "YOLO26 XLarge", "fp16", 112, true),
|
|
800
|
-
ovPrecisionVariant("yolo26x", "objectDetection/yolo26/openvino", "YOLO26 XLarge", "int8", 56, true)
|
|
801
|
-
];
|
|
802
|
-
var FACE_DETECTION_MODELS = [{
|
|
803
|
-
id: "scrfd-2.5g",
|
|
804
|
-
name: "SCRFD 2.5G",
|
|
805
|
-
description: "SCRFD 2.5G — balanced face detection model",
|
|
806
|
-
inputSize: {
|
|
807
|
-
width: 640,
|
|
808
|
-
height: 640
|
|
809
|
-
},
|
|
810
|
-
labels: [{
|
|
811
|
-
id: "face",
|
|
812
|
-
name: "Face"
|
|
813
|
-
}],
|
|
814
|
-
preprocessMode: "letterbox",
|
|
815
|
-
formats: {
|
|
816
|
-
onnx: {
|
|
817
|
-
url: hf("faceDetection/scrfd/onnx/camstack-scrfd-2.5g.onnx"),
|
|
818
|
-
sizeMB: 3.1
|
|
819
|
-
},
|
|
820
|
-
coreml: {
|
|
821
|
-
url: hf("faceDetection/scrfd/coreml/camstack-scrfd-2.5g.mlpackage"),
|
|
822
|
-
sizeMB: 1.7,
|
|
823
|
-
isDirectory: true,
|
|
824
|
-
files: [...MLPACKAGE_FILES],
|
|
825
|
-
runtimes: ["python"]
|
|
826
|
-
},
|
|
827
|
-
openvino: ovFormat(hf("faceDetection/scrfd/openvino/camstack-scrfd-2.5g.xml"), 1.8)
|
|
828
|
-
}
|
|
829
|
-
}, {
|
|
830
|
-
id: "scrypted-yolov9t-face",
|
|
831
|
-
name: "YOLOv9t Face ReLU",
|
|
832
|
-
description: "YOLOv9t face detection — YOLO-based, fast, 320x320 ReLU",
|
|
833
|
-
inputSize: {
|
|
834
|
-
width: 320,
|
|
835
|
-
height: 320
|
|
836
|
-
},
|
|
837
|
-
labels: [{
|
|
838
|
-
id: "face",
|
|
839
|
-
name: "Face"
|
|
840
|
-
}],
|
|
841
|
-
preprocessMode: "letterbox",
|
|
842
|
-
formats: {
|
|
843
|
-
onnx: {
|
|
844
|
-
url: hfScrypted("onnx/scrypted_yolov9t_relu_face/scrypted_yolov9t_relu_face.onnx"),
|
|
845
|
-
sizeMB: 8.1
|
|
846
|
-
},
|
|
847
|
-
coreml: {
|
|
848
|
-
url: hfScrypted("coreml/scrypted_yolov9t_relu_face/scrypted_yolov9t_relu_face.mlpackage"),
|
|
849
|
-
sizeMB: 4,
|
|
850
|
-
isDirectory: true,
|
|
851
|
-
files: [...MLPACKAGE_FILES],
|
|
852
|
-
runtimes: ["python"]
|
|
853
|
-
},
|
|
854
|
-
openvino: ovFormat(hf("faceDetection/scrypted-yolov9-face/openvino/scrypted_yolov9t_relu_face.xml"), 6)
|
|
855
|
-
}
|
|
856
|
-
}];
|
|
857
|
-
var FACE_EMBEDDING_MODELS = [{
|
|
858
|
-
id: "arcface-r100",
|
|
859
|
-
name: "ArcFace R100",
|
|
860
|
-
description: "ArcFace ResNet-100 — high-accuracy face recognition embeddings (512-d)",
|
|
861
|
-
inputSize: {
|
|
862
|
-
width: 112,
|
|
863
|
-
height: 112
|
|
864
|
-
},
|
|
865
|
-
inputLayout: "nhwc",
|
|
866
|
-
labels: [{
|
|
867
|
-
id: "embedding",
|
|
868
|
-
name: "Face Embedding"
|
|
869
|
-
}],
|
|
870
|
-
preprocessMode: "resize",
|
|
871
|
-
faceAlignment: true,
|
|
872
|
-
formats: {
|
|
873
|
-
onnx: {
|
|
874
|
-
url: hf("faceRecognition/arcface/onnx/camstack-arcface-arcface.onnx"),
|
|
875
|
-
sizeMB: 130
|
|
876
|
-
},
|
|
877
|
-
coreml: {
|
|
878
|
-
url: hf("faceRecognition/arcface/coreml/camstack-arcface-r100.mlpackage"),
|
|
879
|
-
sizeMB: 65,
|
|
880
|
-
isDirectory: true,
|
|
881
|
-
files: [...MLPACKAGE_FILES],
|
|
882
|
-
runtimes: ["python"]
|
|
883
|
-
},
|
|
884
|
-
openvino: ovFormat(hf("faceRecognition/arcface/openvino/camstack-arcface-r100.xml"), 65)
|
|
885
|
-
}
|
|
886
|
-
}, {
|
|
887
|
-
id: "inception-resnet-v1",
|
|
888
|
-
name: "Inception ResNet V1",
|
|
889
|
-
description: "FaceNet-style face recognition embeddings (512-d) — hosted on plugin-models HF repo",
|
|
890
|
-
inputSize: {
|
|
891
|
-
width: 160,
|
|
892
|
-
height: 160
|
|
893
|
-
},
|
|
894
|
-
labels: [{
|
|
895
|
-
id: "embedding",
|
|
896
|
-
name: "Face Embedding"
|
|
897
|
-
}],
|
|
898
|
-
preprocessMode: "resize",
|
|
899
|
-
formats: {
|
|
900
|
-
onnx: {
|
|
901
|
-
url: hfScrypted("onnx/inception_resnet_v1/inception_resnet_v1.onnx"),
|
|
902
|
-
sizeMB: 89.6
|
|
903
|
-
},
|
|
904
|
-
coreml: {
|
|
905
|
-
url: hfScrypted("coreml/inception_resnet_v1/inception_resnet_v1.mlpackage"),
|
|
906
|
-
sizeMB: 45,
|
|
907
|
-
isDirectory: true,
|
|
908
|
-
files: [...MLPACKAGE_FILES],
|
|
909
|
-
runtimes: ["python"]
|
|
910
|
-
},
|
|
911
|
-
openvino: ovFormat(hf("faceRecognition/inception-resnet-v1/openvino/camstack-inception-resnet-v1.xml"), 45)
|
|
912
|
-
}
|
|
913
|
-
}];
|
|
914
|
-
var PLATE_DETECTION_MODELS = [{
|
|
915
|
-
id: "yolov8n-plate",
|
|
916
|
-
name: "YOLOv8 Nano — License Plate",
|
|
917
|
-
description: "YOLOv8 Nano fine-tuned for license plate detection",
|
|
918
|
-
inputSize: {
|
|
919
|
-
width: 640,
|
|
920
|
-
height: 640
|
|
921
|
-
},
|
|
922
|
-
labels: [{
|
|
923
|
-
id: "plate",
|
|
924
|
-
name: "License Plate"
|
|
925
|
-
}],
|
|
926
|
-
preprocessMode: "letterbox",
|
|
927
|
-
formats: {
|
|
928
|
-
onnx: {
|
|
929
|
-
url: hf("plateDetection/yolov8-plate/onnx/camstack-yolov8n-plate.onnx"),
|
|
930
|
-
sizeMB: 12
|
|
931
|
-
},
|
|
932
|
-
coreml: {
|
|
933
|
-
url: hf("plateDetection/yolov8-plate/coreml/camstack-yolov8n-plate.mlpackage"),
|
|
934
|
-
sizeMB: 5.9,
|
|
935
|
-
isDirectory: true,
|
|
936
|
-
files: [...MLPACKAGE_FILES],
|
|
937
|
-
runtimes: ["python"]
|
|
938
|
-
},
|
|
939
|
-
openvino: ovFormat(hf("plateDetection/yolov8-plate/openvino/camstack-yolov8n-plate.xml"), 6.1)
|
|
940
|
-
}
|
|
941
|
-
}];
|
|
942
|
-
var PLATE_OCR_MODELS = [{
|
|
943
|
-
id: "vgg-english-g2",
|
|
944
|
-
name: "VGG English G2",
|
|
945
|
-
description: "EasyOCR VGG English G2 — text recognition for license plates and scene text",
|
|
946
|
-
inputSize: {
|
|
947
|
-
width: 384,
|
|
948
|
-
height: 64
|
|
949
|
-
},
|
|
950
|
-
labels: [{
|
|
951
|
-
id: "text",
|
|
952
|
-
name: "Plate Text"
|
|
953
|
-
}],
|
|
954
|
-
preprocessMode: "resize",
|
|
955
|
-
formats: {
|
|
956
|
-
onnx: {
|
|
957
|
-
url: hfScrypted("onnx/vgg_english_g2/vgg_english_g2.onnx"),
|
|
958
|
-
sizeMB: 14.4
|
|
959
|
-
},
|
|
960
|
-
coreml: {
|
|
961
|
-
url: hfScrypted("coreml/vgg_english_g2/vgg_english_g2.mlpackage"),
|
|
962
|
-
sizeMB: 11.2,
|
|
963
|
-
isDirectory: true,
|
|
964
|
-
files: [...MLPACKAGE_FILES],
|
|
965
|
-
runtimes: ["python"]
|
|
966
|
-
},
|
|
967
|
-
openvino: ovFormat(hf("plateRecognition/vgg_english_g2/openvino/vgg_english_g2.xml"), 7.2)
|
|
968
|
-
}
|
|
969
|
-
}];
|
|
970
|
-
var ANIMAL_CLASSIFIER_MODELS = [{
|
|
971
|
-
id: "animals-10",
|
|
972
|
-
name: "Animal Classifier (10)",
|
|
973
|
-
description: "ViT-based animal type classifier — cat, cow, dog, dolphin, eagle, panda, horse, monkey, sheep, spider",
|
|
974
|
-
inputSize: {
|
|
975
|
-
width: 224,
|
|
976
|
-
height: 224
|
|
977
|
-
},
|
|
978
|
-
inputNormalization: "imagenet",
|
|
979
|
-
labels: [{
|
|
980
|
-
id: "animal-type",
|
|
981
|
-
name: "Animal Type"
|
|
982
|
-
}],
|
|
983
|
-
preprocessMode: "resize",
|
|
984
|
-
formats: {
|
|
985
|
-
onnx: {
|
|
986
|
-
url: hf("animalClassification/animals-10/onnx/camstack-animals-10.onnx"),
|
|
987
|
-
sizeMB: 328
|
|
988
|
-
},
|
|
989
|
-
coreml: {
|
|
990
|
-
url: hf("animalClassification/animals-10/coreml/camstack-animals-10.mlpackage"),
|
|
991
|
-
sizeMB: 164,
|
|
992
|
-
isDirectory: true,
|
|
993
|
-
files: [...MLPACKAGE_FILES],
|
|
994
|
-
runtimes: ["python"]
|
|
995
|
-
},
|
|
996
|
-
openvino: ovFormat(hf("animalClassification/animals-10/openvino/camstack-animals-10.xml"), 164)
|
|
997
|
-
}
|
|
998
|
-
}];
|
|
999
|
-
var BIRD_CLASSIFIER_MODELS = [{
|
|
1000
|
-
id: "bird-nabirds-404",
|
|
1001
|
-
name: "NABirds (404 species)",
|
|
1002
|
-
description: "ResNet50 trained on NABirds — 404 North American species",
|
|
1003
|
-
inputSize: {
|
|
1004
|
-
width: 224,
|
|
1005
|
-
height: 224
|
|
1006
|
-
},
|
|
1007
|
-
inputNormalization: "imagenet",
|
|
1008
|
-
labels: [{
|
|
1009
|
-
id: "species",
|
|
1010
|
-
name: "Bird Species"
|
|
1011
|
-
}],
|
|
1012
|
-
preprocessMode: "resize",
|
|
1013
|
-
formats: {
|
|
1014
|
-
onnx: {
|
|
1015
|
-
url: hf("animalClassification/bird-nabirds/onnx/camstack-bird-nabirds-404.onnx"),
|
|
1016
|
-
sizeMB: 93
|
|
1017
|
-
},
|
|
1018
|
-
coreml: {
|
|
1019
|
-
url: hf("animalClassification/bird-nabirds/coreml/camstack-bird-nabirds-404.mlpackage"),
|
|
1020
|
-
sizeMB: 47,
|
|
1021
|
-
isDirectory: true,
|
|
1022
|
-
files: [...MLPACKAGE_FILES],
|
|
1023
|
-
runtimes: ["python"]
|
|
1024
|
-
},
|
|
1025
|
-
openvino: ovFormat(hf("animalClassification/bird-nabirds/openvino/camstack-bird-nabirds-404.xml"), 47)
|
|
1026
|
-
},
|
|
1027
|
-
extraFiles: [{
|
|
1028
|
-
url: hf("animalClassification/bird-nabirds/onnx/camstack-bird-nabirds-404-labels.json"),
|
|
1029
|
-
filename: "camstack-bird-nabirds-404-labels.json",
|
|
1030
|
-
sizeMB: .02
|
|
1031
|
-
}]
|
|
1032
|
-
}];
|
|
1033
|
-
var VEHICLE_CLASSIFIER_MODELS = [{
|
|
1034
|
-
id: "vehicle-type-efficientnet",
|
|
1035
|
-
name: "Vehicle Type (EfficientNet)",
|
|
1036
|
-
description: "EfficientNet-B4 vehicle make/model/year classifier — 8,949 classes from VMMRdb",
|
|
1037
|
-
inputSize: {
|
|
1038
|
-
width: 380,
|
|
1039
|
-
height: 380
|
|
1040
|
-
},
|
|
1041
|
-
inputNormalization: "imagenet",
|
|
1042
|
-
labels: [{
|
|
1043
|
-
id: "vehicle-type",
|
|
1044
|
-
name: "Vehicle Type"
|
|
1045
|
-
}],
|
|
1046
|
-
preprocessMode: "resize",
|
|
1047
|
-
formats: {
|
|
1048
|
-
onnx: {
|
|
1049
|
-
url: hf("vehicleClassification/efficientnet/onnx/camstack-vehicle-type-efficientnet.onnx"),
|
|
1050
|
-
sizeMB: 135
|
|
1051
|
-
},
|
|
1052
|
-
coreml: {
|
|
1053
|
-
url: hf("vehicleClassification/efficientnet/coreml/camstack-vehicle-type-efficientnet.mlpackage"),
|
|
1054
|
-
sizeMB: 10,
|
|
1055
|
-
isDirectory: true,
|
|
1056
|
-
files: [...MLPACKAGE_FILES],
|
|
1057
|
-
runtimes: ["python"]
|
|
1058
|
-
},
|
|
1059
|
-
openvino: ovFormat(hf("vehicleClassification/efficientnet/openvino/camstack-vehicle-type-efficientnet.xml"), 68)
|
|
1060
|
-
},
|
|
1061
|
-
extraFiles: [{
|
|
1062
|
-
url: hf("vehicleClassification/efficientnet/camstack-vehicle-type-labels.json"),
|
|
1063
|
-
filename: "camstack-vehicle-type-labels.json",
|
|
1064
|
-
sizeMB: .2
|
|
1065
|
-
}]
|
|
1066
|
-
}];
|
|
1067
|
-
var SEGMENTATION_REFINER_MODELS = [{
|
|
1068
|
-
id: "u2netp",
|
|
1069
|
-
name: "U2-Net Portable",
|
|
1070
|
-
description: "U2-Net-P — ultra-lightweight salient object segmentation (4.7 MB)",
|
|
1071
|
-
inputSize: {
|
|
1072
|
-
width: 320,
|
|
1073
|
-
height: 320
|
|
1074
|
-
},
|
|
1075
|
-
labels: [{
|
|
1076
|
-
id: "mask",
|
|
1077
|
-
name: "Segmentation Mask"
|
|
1078
|
-
}],
|
|
1079
|
-
preprocessMode: "resize",
|
|
1080
|
-
formats: {
|
|
1081
|
-
onnx: {
|
|
1082
|
-
url: hf("segmentationRefiner/u2netp/onnx/camstack-u2netp.onnx"),
|
|
1083
|
-
sizeMB: 5
|
|
1084
|
-
},
|
|
1085
|
-
coreml: {
|
|
1086
|
-
url: hf("segmentationRefiner/u2netp/coreml/camstack-u2netp.mlpackage"),
|
|
1087
|
-
sizeMB: 3,
|
|
1088
|
-
isDirectory: true,
|
|
1089
|
-
files: [...MLPACKAGE_FILES],
|
|
1090
|
-
runtimes: ["python"]
|
|
1091
|
-
},
|
|
1092
|
-
openvino: ovFormat(hf("segmentationRefiner/u2netp/openvino/camstack-u2netp.xml"), 2.5)
|
|
1093
|
-
}
|
|
1094
|
-
}];
|
|
1095
|
-
var INSTANCE_SEGMENTATION_MODELS = [
|
|
1096
|
-
{
|
|
1097
|
-
id: "yolo26n-seg",
|
|
1098
|
-
name: "YOLO26 Nano Seg",
|
|
1099
|
-
description: "YOLO26 Nano Segmentation — ultra-lightweight instance segmentation with masks",
|
|
1100
|
-
inputSize: {
|
|
1101
|
-
width: 640,
|
|
1102
|
-
height: 640
|
|
1103
|
-
},
|
|
1104
|
-
labels: [],
|
|
1105
|
-
preprocessMode: "letterbox",
|
|
1106
|
-
formats: {
|
|
1107
|
-
onnx: {
|
|
1108
|
-
url: hf("segmentation/yolo26-seg/onnx/camstack-yolo26n-seg.onnx"),
|
|
1109
|
-
sizeMB: 11
|
|
1110
|
-
},
|
|
1111
|
-
coreml: {
|
|
1112
|
-
url: hf("segmentation/yolo26-seg/coreml/camstack-yolo26n-seg.mlpackage"),
|
|
1113
|
-
sizeMB: 6,
|
|
1114
|
-
isDirectory: true,
|
|
1115
|
-
files: [...MLPACKAGE_FILES],
|
|
1116
|
-
runtimes: ["python"]
|
|
1117
|
-
},
|
|
1118
|
-
openvino: ovFormat(hf("segmentation/yolo26-seg/openvino/camstack-yolo26n-seg.xml"), 11)
|
|
1119
|
-
}
|
|
1120
|
-
},
|
|
1121
|
-
{
|
|
1122
|
-
id: "yolo26s-seg",
|
|
1123
|
-
name: "YOLO26 Small Seg",
|
|
1124
|
-
description: "YOLO26 Small Segmentation — balanced instance segmentation",
|
|
1125
|
-
inputSize: {
|
|
1126
|
-
width: 640,
|
|
1127
|
-
height: 640
|
|
1128
|
-
},
|
|
1129
|
-
labels: [],
|
|
1130
|
-
preprocessMode: "letterbox",
|
|
1131
|
-
formats: {
|
|
1132
|
-
onnx: {
|
|
1133
|
-
url: hf("segmentation/yolo26-seg/onnx/camstack-yolo26s-seg.onnx"),
|
|
1134
|
-
sizeMB: 40
|
|
1135
|
-
},
|
|
1136
|
-
coreml: {
|
|
1137
|
-
url: hf("segmentation/yolo26-seg/coreml/camstack-yolo26s-seg.mlpackage"),
|
|
1138
|
-
sizeMB: 20,
|
|
1139
|
-
isDirectory: true,
|
|
1140
|
-
files: [...MLPACKAGE_FILES],
|
|
1141
|
-
runtimes: ["python"]
|
|
1142
|
-
},
|
|
1143
|
-
openvino: ovFormat(hf("segmentation/yolo26-seg/openvino/camstack-yolo26s-seg.xml"), 40)
|
|
1144
|
-
}
|
|
1145
|
-
},
|
|
1146
|
-
{
|
|
1147
|
-
id: "yolo26m-seg",
|
|
1148
|
-
name: "YOLO26 Medium Seg",
|
|
1149
|
-
description: "YOLO26 Medium Segmentation — high-accuracy instance segmentation",
|
|
1150
|
-
inputSize: {
|
|
1151
|
-
width: 640,
|
|
1152
|
-
height: 640
|
|
1153
|
-
},
|
|
1154
|
-
labels: [],
|
|
1155
|
-
preprocessMode: "letterbox",
|
|
1156
|
-
formats: {
|
|
1157
|
-
onnx: {
|
|
1158
|
-
url: hf("segmentation/yolo26-seg/onnx/camstack-yolo26m-seg.onnx"),
|
|
1159
|
-
sizeMB: 90
|
|
1160
|
-
},
|
|
1161
|
-
coreml: {
|
|
1162
|
-
url: hf("segmentation/yolo26-seg/coreml/camstack-yolo26m-seg.mlpackage"),
|
|
1163
|
-
sizeMB: 45,
|
|
1164
|
-
isDirectory: true,
|
|
1165
|
-
files: [...MLPACKAGE_FILES],
|
|
1166
|
-
runtimes: ["python"]
|
|
1167
|
-
},
|
|
1168
|
-
openvino: ovFormat(hf("segmentation/yolo26-seg/openvino/camstack-yolo26m-seg.xml"), 90)
|
|
1169
|
-
}
|
|
1170
|
-
}
|
|
1171
|
-
];
|
|
1172
|
-
var CLIP_EMBEDDING_MODELS = [
|
|
1173
|
-
{
|
|
1174
|
-
id: "mobileclip-s0",
|
|
1175
|
-
name: "MobileCLIP S0",
|
|
1176
|
-
description: "MobileCLIP S0 — Apple ultra-fast CLIP vision encoder, 512-dim, 256×256 (40 MB ONNX)",
|
|
1177
|
-
inputSize: {
|
|
1178
|
-
width: 256,
|
|
1179
|
-
height: 256
|
|
1180
|
-
},
|
|
1181
|
-
labels: [{
|
|
1182
|
-
id: "embedding",
|
|
1183
|
-
name: "CLIP Embedding"
|
|
1184
|
-
}],
|
|
1185
|
-
preprocessMode: "resize",
|
|
1186
|
-
inputNormalization: "none",
|
|
1187
|
-
formats: {
|
|
1188
|
-
onnx: {
|
|
1189
|
-
url: hf("clip/mobileclip-s0/onnx/camstack-mobileclip-s0-vision.onnx"),
|
|
1190
|
-
sizeMB: 40
|
|
1191
|
-
},
|
|
1192
|
-
openvino: ovFormat(hf("clip/mobileclip-s0/openvino/camstack-mobileclip-s0-vision.xml"), 25),
|
|
1193
|
-
coreml: {
|
|
1194
|
-
url: hf("clip/mobileclip-s0/coreml/camstack-mobileclip-s0-vision.mlpackage"),
|
|
1195
|
-
sizeMB: 30,
|
|
1196
|
-
isDirectory: true,
|
|
1197
|
-
files: [...MLPACKAGE_FILES],
|
|
1198
|
-
runtimes: ["python"]
|
|
1199
|
-
}
|
|
1200
|
-
}
|
|
1201
|
-
},
|
|
1202
|
-
{
|
|
1203
|
-
id: "mobileclip-s1",
|
|
1204
|
-
name: "MobileCLIP S1",
|
|
1205
|
-
description: "MobileCLIP S1 — Apple balanced CLIP vision encoder, 512-dim, 256×256 (90 MB ONNX)",
|
|
1206
|
-
inputSize: {
|
|
1207
|
-
width: 256,
|
|
1208
|
-
height: 256
|
|
1209
|
-
},
|
|
1210
|
-
labels: [{
|
|
1211
|
-
id: "embedding",
|
|
1212
|
-
name: "CLIP Embedding"
|
|
1213
|
-
}],
|
|
1214
|
-
preprocessMode: "resize",
|
|
1215
|
-
inputNormalization: "none",
|
|
1216
|
-
formats: {
|
|
1217
|
-
onnx: {
|
|
1218
|
-
url: hf("clip/mobileclip-s1/onnx/camstack-mobileclip-s1-vision.onnx"),
|
|
1219
|
-
sizeMB: 90
|
|
1220
|
-
},
|
|
1221
|
-
openvino: ovFormat(hf("clip/mobileclip-s1/openvino/camstack-mobileclip-s1-vision.xml"), 55),
|
|
1222
|
-
coreml: {
|
|
1223
|
-
url: hf("clip/mobileclip-s1/coreml/camstack-mobileclip-s1-vision.mlpackage"),
|
|
1224
|
-
sizeMB: 65,
|
|
1225
|
-
isDirectory: true,
|
|
1226
|
-
files: [...MLPACKAGE_FILES],
|
|
1227
|
-
runtimes: ["python"]
|
|
1228
|
-
}
|
|
1229
|
-
}
|
|
1230
|
-
},
|
|
1231
|
-
{
|
|
1232
|
-
id: "mobileclip-s2",
|
|
1233
|
-
name: "MobileCLIP S2",
|
|
1234
|
-
description: "MobileCLIP S2 — Apple high-accuracy CLIP vision encoder, 512-dim, 256×256 (150 MB ONNX)",
|
|
1235
|
-
inputSize: {
|
|
1236
|
-
width: 256,
|
|
1237
|
-
height: 256
|
|
1238
|
-
},
|
|
1239
|
-
labels: [{
|
|
1240
|
-
id: "embedding",
|
|
1241
|
-
name: "CLIP Embedding"
|
|
1242
|
-
}],
|
|
1243
|
-
preprocessMode: "resize",
|
|
1244
|
-
inputNormalization: "none",
|
|
1245
|
-
formats: {
|
|
1246
|
-
onnx: {
|
|
1247
|
-
url: hf("clip/mobileclip-s2/onnx/camstack-mobileclip-s2-vision.onnx"),
|
|
1248
|
-
sizeMB: 150
|
|
1249
|
-
},
|
|
1250
|
-
openvino: ovFormat(hf("clip/mobileclip-s2/openvino/camstack-mobileclip-s2-vision.xml"), 90),
|
|
1251
|
-
coreml: {
|
|
1252
|
-
url: hf("clip/mobileclip-s2/coreml/camstack-mobileclip-s2-vision.mlpackage"),
|
|
1253
|
-
sizeMB: 110,
|
|
1254
|
-
isDirectory: true,
|
|
1255
|
-
files: [...MLPACKAGE_FILES],
|
|
1256
|
-
runtimes: ["python"]
|
|
1257
|
-
}
|
|
1258
|
-
}
|
|
1259
|
-
}
|
|
1260
|
-
];
|
|
1261
|
-
var AUDIO_CLASSIFIER_MODELS = [{
|
|
1262
|
-
id: "yamnet-onnx",
|
|
1263
|
-
name: "YAMNet",
|
|
1264
|
-
description: "Google YAMNet — 521-class audio event classifier (3.2 MB ONNX, runs on any platform)",
|
|
1265
|
-
inputSize: {
|
|
1266
|
-
width: 1,
|
|
1267
|
-
height: 16e3
|
|
1268
|
-
},
|
|
1269
|
-
labels: [],
|
|
1270
|
-
preprocessMode: "resize",
|
|
1271
|
-
formats: {
|
|
1272
|
-
onnx: {
|
|
1273
|
-
url: hf("audioClassification/yamnet/onnx/camstack-yamnet.onnx"),
|
|
1274
|
-
sizeMB: 3.2
|
|
1275
|
-
},
|
|
1276
|
-
openvino: ovFormat(hf("audioClassification/yamnet/openvino/camstack-yamnet.xml"), 3.2),
|
|
1277
|
-
coreml: {
|
|
1278
|
-
url: hf("audioClassification/yamnet/onnx/camstack-yamnet.onnx"),
|
|
1279
|
-
sizeMB: 3.2,
|
|
1280
|
-
runtimes: ["python"]
|
|
1281
|
-
}
|
|
1282
|
-
}
|
|
1283
|
-
}, {
|
|
1284
|
-
id: "apple-soundanalysis",
|
|
1285
|
-
name: "Apple SoundAnalysis",
|
|
1286
|
-
description: "macOS built-in — 303 sound categories, Neural Engine accelerated, zero download",
|
|
1287
|
-
inputSize: {
|
|
1288
|
-
width: 1,
|
|
1289
|
-
height: 16e3
|
|
1290
|
-
},
|
|
1291
|
-
labels: [],
|
|
1292
|
-
preprocessMode: "resize",
|
|
1293
|
-
formats: { coreml: {
|
|
1294
|
-
url: "",
|
|
1295
|
-
sizeMB: 0
|
|
1296
|
-
} }
|
|
1297
|
-
}];
|
|
1298
|
-
//#endregion
|
|
1299
|
-
//#region src/detection-pipeline/registry/step-definitions.ts
|
|
1300
|
-
var ANIMAL_10_LABELS = [
|
|
1301
|
-
"cat",
|
|
1302
|
-
"cow",
|
|
1303
|
-
"dog",
|
|
1304
|
-
"dolphin",
|
|
1305
|
-
"eagle",
|
|
1306
|
-
"giant panda",
|
|
1307
|
-
"horse",
|
|
1308
|
-
"monkey",
|
|
1309
|
-
"sheep",
|
|
1310
|
-
"spider"
|
|
1311
|
-
];
|
|
1312
|
-
var VGG_G2_CHARSET = ["", ..."0123456789!\"#$%&'()*+,-./:;<=>?@[\\]^_`{|}~ ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz".split("")];
|
|
1313
|
-
var PipelineStepBase = class {
|
|
1314
|
-
definition;
|
|
1315
|
-
constructor(definition) {
|
|
1316
|
-
this.definition = definition;
|
|
1317
|
-
}
|
|
1318
|
-
getConfigSchema() {
|
|
1319
|
-
return [];
|
|
1320
|
-
}
|
|
1321
|
-
};
|
|
1322
|
-
var ObjectDetectionStep = class {
|
|
1323
|
-
definition = {
|
|
1324
|
-
id: "object-detection",
|
|
1325
|
-
name: "Object Detection",
|
|
1326
|
-
slot: "detector",
|
|
1327
|
-
postprocessor: "yolo",
|
|
1328
|
-
extractMode: "full-frame",
|
|
1329
|
-
inputClasses: null,
|
|
1330
|
-
outputClasses: [
|
|
1331
|
-
"person",
|
|
1332
|
-
"vehicle",
|
|
1333
|
-
"animal"
|
|
1334
|
-
],
|
|
1335
|
-
models: [...OBJECT_DETECTION_MODELS],
|
|
1336
|
-
defaultModelId: "yolo26n",
|
|
1337
|
-
defaultConfidence: .5,
|
|
1338
|
-
labels: COCO_80_LABELS.map((l) => l.id),
|
|
1339
|
-
classMap: COCO_TO_MACRO
|
|
1340
|
-
};
|
|
1341
|
-
getConfigSchema() {
|
|
1342
|
-
return [
|
|
1343
|
-
{
|
|
1344
|
-
type: "multiselect",
|
|
1345
|
-
key: "enabledMacroClasses",
|
|
1346
|
-
label: "Enabled Classes",
|
|
1347
|
-
description: "Which macro classes to detect (empty = all three).",
|
|
1348
|
-
options: [
|
|
1349
|
-
{
|
|
1350
|
-
value: "person",
|
|
1351
|
-
label: "Person"
|
|
1352
|
-
},
|
|
1353
|
-
{
|
|
1354
|
-
value: "vehicle",
|
|
1355
|
-
label: "Vehicle"
|
|
1356
|
-
},
|
|
1357
|
-
{
|
|
1358
|
-
value: "animal",
|
|
1359
|
-
label: "Animal"
|
|
1360
|
-
}
|
|
1361
|
-
],
|
|
1362
|
-
default: [
|
|
1363
|
-
"person",
|
|
1364
|
-
"vehicle",
|
|
1365
|
-
"animal"
|
|
1366
|
-
]
|
|
1367
|
-
},
|
|
1368
|
-
{
|
|
1369
|
-
type: "slider",
|
|
1370
|
-
key: "minConfidencePerson",
|
|
1371
|
-
label: "Min Confidence — Person",
|
|
1372
|
-
description: "Drop person predictions below this score.",
|
|
1373
|
-
min: 0,
|
|
1374
|
-
max: 1,
|
|
1375
|
-
step: .05,
|
|
1376
|
-
default: .5,
|
|
1377
|
-
showValue: true
|
|
1378
|
-
},
|
|
1379
|
-
{
|
|
1380
|
-
type: "slider",
|
|
1381
|
-
key: "minConfidenceVehicle",
|
|
1382
|
-
label: "Min Confidence — Vehicle",
|
|
1383
|
-
description: "Drop vehicle predictions below this score.",
|
|
1384
|
-
min: 0,
|
|
1385
|
-
max: 1,
|
|
1386
|
-
step: .05,
|
|
1387
|
-
default: .5,
|
|
1388
|
-
showValue: true
|
|
1389
|
-
},
|
|
1390
|
-
{
|
|
1391
|
-
type: "slider",
|
|
1392
|
-
key: "minConfidenceAnimal",
|
|
1393
|
-
label: "Min Confidence — Animal",
|
|
1394
|
-
description: "Drop animal predictions below this score.",
|
|
1395
|
-
min: 0,
|
|
1396
|
-
max: 1,
|
|
1397
|
-
step: .05,
|
|
1398
|
-
default: .5,
|
|
1399
|
-
showValue: true
|
|
1400
|
-
}
|
|
1401
|
-
];
|
|
1402
|
-
}
|
|
1403
|
-
};
|
|
1404
|
-
var AudioClassifierStep = class {
|
|
1405
|
-
definition = {
|
|
1406
|
-
id: "audio-classifier",
|
|
1407
|
-
name: "Audio Classifier",
|
|
1408
|
-
slot: "audio-classifier",
|
|
1409
|
-
postprocessor: "yamnet",
|
|
1410
|
-
extractMode: "full-frame",
|
|
1411
|
-
inputClasses: null,
|
|
1412
|
-
outputClasses: ["audio"],
|
|
1413
|
-
models: [...AUDIO_CLASSIFIER_MODELS],
|
|
1414
|
-
defaultModelId: "yamnet-onnx",
|
|
1415
|
-
defaultModelIdByFormat: { coreml: "apple-soundanalysis" },
|
|
1416
|
-
enabledByDefault: true,
|
|
1417
|
-
defaultConfidence: .3
|
|
1418
|
-
};
|
|
1419
|
-
getConfigSchema() {
|
|
1420
|
-
return [{
|
|
1421
|
-
type: "multiselect",
|
|
1422
|
-
key: "enabledAudioClasses",
|
|
1423
|
-
label: "Audio Classes",
|
|
1424
|
-
description: "Which audio macro classes to surface (empty = all).",
|
|
1425
|
-
options: AUDIO_MACRO_LABELS.map((l) => ({
|
|
1426
|
-
value: l.id,
|
|
1427
|
-
label: `${l.icon ?? ""} ${l.name}`.trim()
|
|
1428
|
-
})),
|
|
1429
|
-
default: [
|
|
1430
|
-
"scream",
|
|
1431
|
-
"crying",
|
|
1432
|
-
"dog",
|
|
1433
|
-
"alarm",
|
|
1434
|
-
"doorbell",
|
|
1435
|
-
"glass_breaking",
|
|
1436
|
-
"gunshot",
|
|
1437
|
-
"siren",
|
|
1438
|
-
"footsteps"
|
|
1439
|
-
]
|
|
1440
|
-
}, {
|
|
1441
|
-
type: "slider",
|
|
1442
|
-
key: "minConfidence",
|
|
1443
|
-
label: "Min Confidence",
|
|
1444
|
-
description: "Drop predictions whose score is below this threshold.",
|
|
1445
|
-
default: .3,
|
|
1446
|
-
min: 0,
|
|
1447
|
-
max: 1,
|
|
1448
|
-
step: .05,
|
|
1449
|
-
showValue: true
|
|
1450
|
-
}];
|
|
1451
|
-
}
|
|
1452
|
-
};
|
|
1453
|
-
var AnimalClassifierStep = class {
|
|
1454
|
-
definition = {
|
|
1455
|
-
id: "animal-classifier",
|
|
1456
|
-
name: "Animal Classifier",
|
|
1457
|
-
slot: "classifier",
|
|
1458
|
-
postprocessor: "softmax",
|
|
1459
|
-
extractMode: "crop-roi",
|
|
1460
|
-
inputClasses: ["animal"],
|
|
1461
|
-
outputClasses: ["animal-type"],
|
|
1462
|
-
models: [...ANIMAL_CLASSIFIER_MODELS],
|
|
1463
|
-
defaultModelId: "animals-10",
|
|
1464
|
-
enabledByDefault: false,
|
|
1465
|
-
defaultConfidence: .2,
|
|
1466
|
-
labels: [...ANIMAL_10_LABELS]
|
|
1467
|
-
};
|
|
1468
|
-
getConfigSchema() {
|
|
1469
|
-
return [{
|
|
1470
|
-
type: "multiselect",
|
|
1471
|
-
key: "enabledClasses",
|
|
1472
|
-
label: "Species",
|
|
1473
|
-
description: "Which animal species to keep (empty = all). The classifier still scores every label but only the selected classes are returned on the detection.",
|
|
1474
|
-
options: ANIMAL_10_LABELS.map((l) => ({
|
|
1475
|
-
value: l,
|
|
1476
|
-
label: l
|
|
1477
|
-
})),
|
|
1478
|
-
default: []
|
|
1479
|
-
}, {
|
|
1480
|
-
type: "slider",
|
|
1481
|
-
key: "minConfidence",
|
|
1482
|
-
label: "Min Confidence",
|
|
1483
|
-
description: "Drop classifications whose score is below this threshold.",
|
|
1484
|
-
default: this.definition.defaultConfidence,
|
|
1485
|
-
min: 0,
|
|
1486
|
-
max: 1,
|
|
1487
|
-
step: .05,
|
|
1488
|
-
showValue: true
|
|
1489
|
-
}];
|
|
1490
|
-
}
|
|
1491
|
-
};
|
|
1492
|
-
var ClassifierWithMinConfidence = class extends PipelineStepBase {
|
|
1493
|
-
getConfigSchema() {
|
|
1494
|
-
return [{
|
|
1495
|
-
type: "slider",
|
|
1496
|
-
key: "minConfidence",
|
|
1497
|
-
label: "Min Confidence",
|
|
1498
|
-
description: "Drop classifications whose score is below this threshold.",
|
|
1499
|
-
default: this.definition.defaultConfidence,
|
|
1500
|
-
min: 0,
|
|
1501
|
-
max: 1,
|
|
1502
|
-
step: .05,
|
|
1503
|
-
showValue: true
|
|
1504
|
-
}];
|
|
1505
|
-
}
|
|
1506
|
-
};
|
|
1507
|
-
var STEP_FACE_DETECTION = new PipelineStepBase({
|
|
1508
|
-
id: "face-detection",
|
|
1509
|
-
name: "Face Detection",
|
|
1510
|
-
slot: "cropper",
|
|
1511
|
-
postprocessor: "scrfd",
|
|
1512
|
-
extractMode: "crop-roi",
|
|
1513
|
-
inputClasses: ["person"],
|
|
1514
|
-
outputClasses: ["face"],
|
|
1515
|
-
models: [...FACE_DETECTION_MODELS],
|
|
1516
|
-
defaultModelId: "scrfd-2.5g",
|
|
1517
|
-
defaultConfidence: .5,
|
|
1518
|
-
defaultMinParentScore: .7
|
|
1519
|
-
});
|
|
1520
|
-
var STEP_FACE_EMBEDDING = new PipelineStepBase({
|
|
1521
|
-
id: "face-embedding",
|
|
1522
|
-
name: "Face Embedding",
|
|
1523
|
-
slot: "classifier",
|
|
1524
|
-
postprocessor: "arcface",
|
|
1525
|
-
extractMode: "crop-roi",
|
|
1526
|
-
inputClasses: ["face"],
|
|
1527
|
-
outputClasses: ["identity"],
|
|
1528
|
-
models: [...FACE_EMBEDDING_MODELS],
|
|
1529
|
-
defaultModelId: "arcface-r100",
|
|
1530
|
-
defaultConfidence: 0
|
|
1531
|
-
});
|
|
1532
|
-
var STEP_CLIP_EMBEDDING = new PipelineStepBase({
|
|
1533
|
-
id: "clip-embedding",
|
|
1534
|
-
name: "CLIP Embedding",
|
|
1535
|
-
slot: "classifier",
|
|
1536
|
-
postprocessor: "clip",
|
|
1537
|
-
extractMode: "crop-roi",
|
|
1538
|
-
inputClasses: [
|
|
1539
|
-
"person",
|
|
1540
|
-
"vehicle",
|
|
1541
|
-
"animal"
|
|
1542
|
-
],
|
|
1543
|
-
outputClasses: ["embedding"],
|
|
1544
|
-
models: [...CLIP_EMBEDDING_MODELS],
|
|
1545
|
-
defaultModelId: "mobileclip-s1",
|
|
1546
|
-
enabledByDefault: true,
|
|
1547
|
-
backfillIntoExistingOverrides: true,
|
|
1548
|
-
defaultConfidence: 0,
|
|
1549
|
-
defaultMinParentScore: .4,
|
|
1550
|
-
group: "CLIP"
|
|
1551
|
-
});
|
|
1552
|
-
var STEP_PLATE_DETECTION = new PipelineStepBase({
|
|
1553
|
-
id: "plate-detection",
|
|
1554
|
-
name: "Plate Detection",
|
|
1555
|
-
slot: "cropper",
|
|
1556
|
-
postprocessor: "yolo",
|
|
1557
|
-
extractMode: "crop-roi",
|
|
1558
|
-
inputClasses: ["vehicle"],
|
|
1559
|
-
outputClasses: ["plate"],
|
|
1560
|
-
models: [...PLATE_DETECTION_MODELS],
|
|
1561
|
-
defaultModelId: "yolov8n-plate",
|
|
1562
|
-
defaultConfidence: .5,
|
|
1563
|
-
defaultMinParentScore: .7,
|
|
1564
|
-
labels: ["plate"]
|
|
1565
|
-
});
|
|
1566
|
-
var STEP_PLATE_OCR = new PipelineStepBase({
|
|
1567
|
-
id: "plate-ocr",
|
|
1568
|
-
name: "Plate OCR",
|
|
1569
|
-
slot: "classifier",
|
|
1570
|
-
postprocessor: "ctc",
|
|
1571
|
-
extractMode: "crop-roi",
|
|
1572
|
-
inputClasses: ["plate"],
|
|
1573
|
-
outputClasses: ["plate-text"],
|
|
1574
|
-
models: [...PLATE_OCR_MODELS],
|
|
1575
|
-
defaultModelId: "vgg-english-g2",
|
|
1576
|
-
defaultConfidence: 0,
|
|
1577
|
-
charset: [...VGG_G2_CHARSET]
|
|
1578
|
-
});
|
|
1579
|
-
var STEP_BIRD_CLASSIFIER = new ClassifierWithMinConfidence({
|
|
1580
|
-
id: "bird-classifier",
|
|
1581
|
-
name: "Bird Classifier",
|
|
1582
|
-
slot: "classifier",
|
|
1583
|
-
postprocessor: "softmax",
|
|
1584
|
-
extractMode: "crop-roi",
|
|
1585
|
-
inputClasses: ["animal"],
|
|
1586
|
-
outputClasses: ["species"],
|
|
1587
|
-
models: [...BIRD_CLASSIFIER_MODELS],
|
|
1588
|
-
defaultModelId: "bird-nabirds-404",
|
|
1589
|
-
enabledByDefault: false,
|
|
1590
|
-
defaultConfidence: .2
|
|
1591
|
-
});
|
|
1592
|
-
var STEP_VEHICLE_CLASSIFIER = new ClassifierWithMinConfidence({
|
|
1593
|
-
id: "vehicle-classifier",
|
|
1594
|
-
name: "Vehicle Classifier",
|
|
1595
|
-
slot: "classifier",
|
|
1596
|
-
postprocessor: "softmax",
|
|
1597
|
-
extractMode: "crop-roi",
|
|
1598
|
-
inputClasses: ["vehicle"],
|
|
1599
|
-
outputClasses: ["vehicle-type"],
|
|
1600
|
-
models: [...VEHICLE_CLASSIFIER_MODELS],
|
|
1601
|
-
defaultModelId: "vehicle-type-efficientnet",
|
|
1602
|
-
enabledByDefault: false,
|
|
1603
|
-
defaultConfidence: .3
|
|
1604
|
-
});
|
|
1605
|
-
var STEP_SEGMENTATION = new PipelineStepBase({
|
|
1606
|
-
id: "segmentation",
|
|
1607
|
-
name: "Segmentation",
|
|
1608
|
-
slot: "refiner",
|
|
1609
|
-
postprocessor: "saliency",
|
|
1610
|
-
extractMode: "crop-roi",
|
|
1611
|
-
inputClasses: null,
|
|
1612
|
-
outputClasses: ["mask"],
|
|
1613
|
-
models: [...SEGMENTATION_REFINER_MODELS],
|
|
1614
|
-
defaultModelId: "u2netp",
|
|
1615
|
-
enabledByDefault: false,
|
|
1616
|
-
defaultConfidence: 0,
|
|
1617
|
-
group: "Segmentation"
|
|
1618
|
-
});
|
|
1619
|
-
new PipelineStepBase({
|
|
1620
|
-
id: "instance-segmentation",
|
|
1621
|
-
name: "Instance Segmentation",
|
|
1622
|
-
slot: "refiner",
|
|
1623
|
-
postprocessor: "yolo-seg",
|
|
1624
|
-
extractMode: "full-frame",
|
|
1625
|
-
inputClasses: null,
|
|
1626
|
-
outputClasses: ["mask"],
|
|
1627
|
-
models: [...INSTANCE_SEGMENTATION_MODELS],
|
|
1628
|
-
defaultModelId: "yolo26n-seg",
|
|
1629
|
-
enabledByDefault: false,
|
|
1630
|
-
defaultConfidence: .25,
|
|
1631
|
-
labels: COCO_80_LABELS.map((l) => l.id),
|
|
1632
|
-
classMap: COCO_TO_MACRO,
|
|
1633
|
-
group: "Segmentation"
|
|
1634
|
-
});
|
|
1635
|
-
var STEP_OBJECT_DETECTION_INSTANCE = new ObjectDetectionStep();
|
|
1636
|
-
var STEP_AUDIO_CLASSIFIER_INSTANCE = new AudioClassifierStep();
|
|
1637
|
-
var ALL_PIPELINE_STEPS = [
|
|
1638
|
-
STEP_OBJECT_DETECTION_INSTANCE,
|
|
1639
|
-
STEP_FACE_DETECTION,
|
|
1640
|
-
STEP_FACE_EMBEDDING,
|
|
1641
|
-
STEP_CLIP_EMBEDDING,
|
|
1642
|
-
STEP_PLATE_DETECTION,
|
|
1643
|
-
STEP_PLATE_OCR,
|
|
1644
|
-
new AnimalClassifierStep(),
|
|
1645
|
-
STEP_BIRD_CLASSIFIER,
|
|
1646
|
-
STEP_VEHICLE_CLASSIFIER,
|
|
1647
|
-
STEP_SEGMENTATION,
|
|
1648
|
-
STEP_AUDIO_CLASSIFIER_INSTANCE
|
|
1649
|
-
];
|
|
1650
|
-
/** Compat: flat array of StepDefinition for existing consumers */
|
|
1651
|
-
var ALL_STEPS = ALL_PIPELINE_STEPS.map((s) => s.definition);
|
|
1652
|
-
var STEP_MAP = new Map(ALL_PIPELINE_STEPS.map((s) => [s.definition.id, s]));
|
|
1653
|
-
/**
|
|
1654
|
-
* Look up a pipeline step by ID (class instance with getConfigSchema).
|
|
1655
|
-
* @throws if the step ID is not registered.
|
|
1656
|
-
*/
|
|
1657
|
-
function getStep(stepId) {
|
|
1658
|
-
const step = STEP_MAP.get(stepId);
|
|
1659
|
-
if (!step) throw new Error(`Unknown pipeline step: "${stepId}". Available: ${ALL_PIPELINE_STEPS.map((s) => s.definition.id).join(", ")}`);
|
|
1660
|
-
return step;
|
|
1661
|
-
}
|
|
1662
|
-
/**
|
|
1663
|
-
* Look up a step definition by ID (compat shortcut).
|
|
1664
|
-
* @throws if the step ID is not registered.
|
|
1665
|
-
*/
|
|
1666
|
-
function getStepDefinition(stepId) {
|
|
1667
|
-
return getStep(stepId).definition;
|
|
1668
|
-
}
|
|
1669
|
-
/**
|
|
1670
|
-
* Get the default model ID for a step given the current model format.
|
|
1671
|
-
*
|
|
1672
|
-
* Resolution order:
|
|
1673
|
-
* 1. `def.defaultModelIdByFormat?.[format]` — the step's declared
|
|
1674
|
-
* per-format preference — if that model exists in `def.models` AND
|
|
1675
|
-
* has a build for `format`.
|
|
1676
|
-
* 2. `def.defaultModelId` — the step's plain declared default — if it
|
|
1677
|
-
* exists in `def.models` AND has a build for `format`.
|
|
1678
|
-
* 3. The smallest-by-size model among those with a `format` build
|
|
1679
|
-
* (legacy fallback, preserved for steps/formats with no declared
|
|
1680
|
-
* preference reachable).
|
|
1681
|
-
* 4. `def.defaultModelId` unchanged, when ZERO models have a `format`
|
|
1682
|
-
* build — an unloadable case flagged elsewhere, not resolved here.
|
|
1683
|
-
*/
|
|
1684
|
-
function getDefaultModelForFormat(stepId, format) {
|
|
1685
|
-
const def = getStepDefinition(stepId);
|
|
1686
|
-
const hasFormatBuild = (modelId) => def.models.find((m) => m.id === modelId)?.formats[format] !== void 0;
|
|
1687
|
-
const declaredForFormat = def.defaultModelIdByFormat?.[format];
|
|
1688
|
-
if (declaredForFormat !== void 0 && hasFormatBuild(declaredForFormat)) return declaredForFormat;
|
|
1689
|
-
if (hasFormatBuild(def.defaultModelId)) return def.defaultModelId;
|
|
1690
|
-
const available = def.models.filter((m) => m.formats[format] && m.legacy !== true);
|
|
1691
|
-
if (available.length === 0) return def.defaultModelId;
|
|
1692
|
-
return [...available].toSorted((a, b) => {
|
|
1693
|
-
return (a.formats[format]?.sizeMB ?? Infinity) - (b.formats[format]?.sizeMB ?? Infinity);
|
|
1694
|
-
})[0].id;
|
|
1695
|
-
}
|
|
1696
|
-
/**
|
|
1697
|
-
* The single node-local authority for turning a CHOSEN model (or none) into a
|
|
1698
|
-
* concrete model that has a build for this node's engine `format`. Fuses the
|
|
1699
|
-
* two previously-duplicated decisions (orchestrator `pickDefaultModelId` +
|
|
1700
|
-
* provider `substituteIncompatibleModels`):
|
|
1701
|
-
*
|
|
1702
|
-
* - chosen === null → the addon's default model for `format`.
|
|
1703
|
-
* - chosen is a catalog model → keep it if it has a `format` build; else
|
|
1704
|
-
* the addon's default model for `format`
|
|
1705
|
-
* (a substitution — the caller surfaces it).
|
|
1706
|
-
* - chosen is NOT in the catalog (custom/unknown) → returned unchanged; the
|
|
1707
|
-
* operator owns a custom model's format compat.
|
|
1708
|
-
* - unknown stepId → chosen returned unchanged (never throws).
|
|
1709
|
-
*/
|
|
1710
|
-
function resolveModelForFormat(stepId, chosenModelId, format) {
|
|
1711
|
-
let def;
|
|
1712
|
-
try {
|
|
1713
|
-
def = getStepDefinition(stepId);
|
|
1714
|
-
} catch {
|
|
1715
|
-
return chosenModelId ?? "";
|
|
1716
|
-
}
|
|
1717
|
-
if (chosenModelId === null) return getDefaultModelForFormat(stepId, format);
|
|
1718
|
-
const entry = def.models.find((m) => m.id === chosenModelId);
|
|
1719
|
-
if (!entry) return chosenModelId;
|
|
1720
|
-
if (entry.formats[format]) return chosenModelId;
|
|
1721
|
-
return getDefaultModelForFormat(stepId, format);
|
|
1722
|
-
}
|
|
1723
|
-
//#endregion
|
|
1724
294
|
//#region src/detection-pipeline/default-detection-model.ts
|
|
1725
295
|
/** The object-detection step id — the only slot this resolver applies to. */
|
|
1726
296
|
var OBJECT_DETECTION_STEP_ID = "object-detection";
|
|
@@ -3715,6 +2285,110 @@ function bboxShortSide(bbox) {
|
|
|
3715
2285
|
return Math.min(bbox[2] - bbox[0], bbox[3] - bbox[1]);
|
|
3716
2286
|
}
|
|
3717
2287
|
/**
|
|
2288
|
+
* Padded region (clamped to the frame) around a face bbox — the area a native
|
|
2289
|
+
* crop must cover so the warp has room around a tilted / edge face. Mirrors the
|
|
2290
|
+
* padding {@link alignFaceCrop} applies internally (`padRatio` default 0.4), so
|
|
2291
|
+
* the native and fallback paths sample the same neighbourhood.
|
|
2292
|
+
*/
|
|
2293
|
+
function paddedFaceRegion(faceBbox, imageWidth, imageHeight, padRatio = .4) {
|
|
2294
|
+
const bw = faceBbox[2] - faceBbox[0];
|
|
2295
|
+
const bh = faceBbox[3] - faceBbox[1];
|
|
2296
|
+
const padX = bw * padRatio;
|
|
2297
|
+
const padY = bh * padRatio;
|
|
2298
|
+
return {
|
|
2299
|
+
left: Math.max(0, Math.floor(faceBbox[0] - padX)),
|
|
2300
|
+
top: Math.max(0, Math.floor(faceBbox[1] - padY)),
|
|
2301
|
+
right: Math.min(imageWidth, Math.ceil(faceBbox[2] + padX)),
|
|
2302
|
+
bottom: Math.min(imageHeight, Math.ceil(faceBbox[3] + padY))
|
|
2303
|
+
};
|
|
2304
|
+
}
|
|
2305
|
+
/**
|
|
2306
|
+
* Warp a raw RGB crop of `region` (native-resolution pixels of exactly that
|
|
2307
|
+
* detection-space region) so `landmarksImageSpace` land on the ArcFace
|
|
2308
|
+
* template. Landmarks are mapped from image space into the crop's local pixel
|
|
2309
|
+
* space by the per-axis region→crop scale (handles `maxWidth` downscale of the
|
|
2310
|
+
* native crop). Channels default to RGB (3) — the format the decode worker's
|
|
2311
|
+
* native surface produces.
|
|
2312
|
+
*/
|
|
2313
|
+
async function alignFaceCropFromRgb(rgb, region, landmarksImageSpace, opts = {}) {
|
|
2314
|
+
const outSize = opts.outSize ?? 112;
|
|
2315
|
+
const quality = opts.quality ?? 95;
|
|
2316
|
+
const channels = opts.channels ?? 3;
|
|
2317
|
+
const regionW = region.right - region.left;
|
|
2318
|
+
const regionH = region.bottom - region.top;
|
|
2319
|
+
if (regionW < 1 || regionH < 1) throw new Error(`alignFaceCropFromRgb: degenerate region ${regionW}x${regionH}`);
|
|
2320
|
+
if (rgb.width < 2 || rgb.height < 2) throw new Error(`alignFaceCropFromRgb: degenerate crop ${rgb.width}x${rgb.height}`);
|
|
2321
|
+
const scaleX = rgb.width / regionW;
|
|
2322
|
+
const scaleY = rgb.height / regionH;
|
|
2323
|
+
const localLandmarks = landmarksImageSpace.map((l) => ({
|
|
2324
|
+
x: (l.x - region.left) * scaleX,
|
|
2325
|
+
y: (l.y - region.top) * scaleY
|
|
2326
|
+
}));
|
|
2327
|
+
const aligned = warpToTemplate({
|
|
2328
|
+
data: rgb.bytes,
|
|
2329
|
+
width: rgb.width,
|
|
2330
|
+
height: rgb.height,
|
|
2331
|
+
channels
|
|
2332
|
+
}, localLandmarks, opts.template ?? ARCFACE_TEMPLATE_112, outSize);
|
|
2333
|
+
return {
|
|
2334
|
+
jpeg: await sharp(Buffer.from(aligned.data), { raw: {
|
|
2335
|
+
width: aligned.width,
|
|
2336
|
+
height: aligned.height,
|
|
2337
|
+
channels: aligned.channels
|
|
2338
|
+
} }).jpeg({ quality }).toBuffer(),
|
|
2339
|
+
width: aligned.width,
|
|
2340
|
+
height: aligned.height
|
|
2341
|
+
};
|
|
2342
|
+
}
|
|
2343
|
+
/**
|
|
2344
|
+
* Produce a landmark-aligned face crop, preferring a NATIVE-resolution crop
|
|
2345
|
+
* (design B — the quality win) and FALLING BACK to today's downscaled
|
|
2346
|
+
* full-frame alignment on any native miss/error (never worse than before).
|
|
2347
|
+
*
|
|
2348
|
+
* The native path requests the padded face region from `nativeCropProvider`
|
|
2349
|
+
* (a best-effort local cap round-trip bound to this frame), warps the native
|
|
2350
|
+
* pixels; the fallback path re-uses {@link alignFaceCrop} on the downscaled
|
|
2351
|
+
* JPEG. ArcFace input is a fixed `outSize`×`outSize` either way — only the crop
|
|
2352
|
+
* SOURCE changes, so model cost is unchanged.
|
|
2353
|
+
*/
|
|
2354
|
+
async function buildAlignedFaceCrop(args) {
|
|
2355
|
+
if (args.nativeCropProvider) try {
|
|
2356
|
+
const region = paddedFaceRegion(args.faceBbox, args.imageWidth, args.imageHeight);
|
|
2357
|
+
const roi = {
|
|
2358
|
+
x: region.left / args.imageWidth,
|
|
2359
|
+
y: region.top / args.imageHeight,
|
|
2360
|
+
w: (region.right - region.left) / args.imageWidth,
|
|
2361
|
+
h: (region.bottom - region.top) / args.imageHeight
|
|
2362
|
+
};
|
|
2363
|
+
const native = await args.nativeCropProvider(roi, args.nativeMaxWidth);
|
|
2364
|
+
if (native && native.width >= 2 && native.height >= 2 && native.bytes.length >= native.width * native.height * 3) return {
|
|
2365
|
+
...await alignFaceCropFromRgb(native, region, args.landmarksImageSpace, { outSize: args.outSize }),
|
|
2366
|
+
source: "native",
|
|
2367
|
+
nativeFaceShortSidePx: nativeFaceShortSide(args.faceBbox, region, native)
|
|
2368
|
+
};
|
|
2369
|
+
} catch {}
|
|
2370
|
+
return {
|
|
2371
|
+
...await alignFaceCrop(args.fullFrameJpeg, args.faceBbox, args.landmarksImageSpace, args.imageWidth, args.imageHeight, { outSize: args.outSize }),
|
|
2372
|
+
source: "fallback"
|
|
2373
|
+
};
|
|
2374
|
+
}
|
|
2375
|
+
/**
|
|
2376
|
+
* Short side (px) of `faceBbox` (in detection space) re-measured in the NATIVE
|
|
2377
|
+
* crop's resolution. The native crop covers `region` (detection-space padded
|
|
2378
|
+
* region) rendered at `native.width × native.height` native pixels, so the
|
|
2379
|
+
* per-axis scale `native / region` maps the detection face size to its native
|
|
2380
|
+
* size. This is the size an embedder actually sees, so the gate uses it.
|
|
2381
|
+
*/
|
|
2382
|
+
function nativeFaceShortSide(faceBbox, region, native) {
|
|
2383
|
+
const regionW = Math.max(1, region.right - region.left);
|
|
2384
|
+
const regionH = Math.max(1, region.bottom - region.top);
|
|
2385
|
+
const scaleX = native.width / regionW;
|
|
2386
|
+
const scaleY = native.height / regionH;
|
|
2387
|
+
const faceW = (faceBbox[2] - faceBbox[0]) * scaleX;
|
|
2388
|
+
const faceH = (faceBbox[3] - faceBbox[1]) * scaleY;
|
|
2389
|
+
return Math.min(faceW, faceH);
|
|
2390
|
+
}
|
|
2391
|
+
/**
|
|
3718
2392
|
* Produce an aligned 112x112 JPEG of a face from the full-frame JPEG and the
|
|
3719
2393
|
* face's 5 landmarks (in absolute image-space coordinates). Decodes a padded
|
|
3720
2394
|
* region around the face bbox to raw RGB, warps to the ArcFace template, and
|
|
@@ -3942,7 +2616,8 @@ function buildFrameResult(input) {
|
|
|
3942
2616
|
...embedding !== void 0 ? {
|
|
3943
2617
|
embedding,
|
|
3944
2618
|
...embeddingModelId !== void 0 ? { embeddingModelId } : {}
|
|
3945
|
-
} : {}
|
|
2619
|
+
} : {},
|
|
2620
|
+
...m.faceAlignedCrop !== void 0 ? { faceAlignedCrop: m.faceAlignedCrop } : {}
|
|
3946
2621
|
};
|
|
3947
2622
|
if (input.debug) {
|
|
3948
2623
|
const cleanDebug = pruneUndefined({
|
|
@@ -4023,7 +2698,8 @@ function toMutableRootDetection(det, rootStep, idGen, rootStepLatencyMs) {
|
|
|
4023
2698
|
mask: det.mask,
|
|
4024
2699
|
maskWidth: det.maskWidth,
|
|
4025
2700
|
maskHeight: det.maskHeight
|
|
4026
|
-
} : {}
|
|
2701
|
+
} : {},
|
|
2702
|
+
...det.landmarks !== void 0 ? { landmarks: det.landmarks } : {}
|
|
4027
2703
|
};
|
|
4028
2704
|
}
|
|
4029
2705
|
//#endregion
|
|
@@ -4141,6 +2817,33 @@ function formatTopResult(output) {
|
|
|
4141
2817
|
* resolution — a deeper change tracked separately.
|
|
4142
2818
|
*/
|
|
4143
2819
|
var DEFAULT_MIN_FACE_SIZE_PX = 28;
|
|
2820
|
+
/**
|
|
2821
|
+
* Cheap pre-native floor (px, detection space): below this a face is treated as
|
|
2822
|
+
* genuinely degenerate — SCRFD landmarks are unreliable and even a large native
|
|
2823
|
+
* upscale can't rescue it — so it is skipped WITHOUT paying a native-crop
|
|
2824
|
+
* round-trip. Above the floor, the real {@link DEFAULT_MIN_FACE_SIZE_PX} gate is
|
|
2825
|
+
* evaluated on the NATIVE face size (see the face branch of `executeChildren`).
|
|
2826
|
+
*/
|
|
2827
|
+
var MIN_FACE_DETECTION_FLOOR_PX = 6;
|
|
2828
|
+
/**
|
|
2829
|
+
* Cap (px) on the width of the NATIVE face crop fetched for embedding (design
|
|
2830
|
+
* B). The arcface input is a fixed 112² warp, so a wider native crop buys no
|
|
2831
|
+
* embedding quality past a modest margin but does inflate the per-face transport
|
|
2832
|
+
* of the best-effort native-crop round-trip. 320px keeps a distant face well
|
|
2833
|
+
* above the downscaled detection resolution while bounding the crop payload.
|
|
2834
|
+
*/
|
|
2835
|
+
var NATIVE_FACE_CROP_MAX_WIDTH = 320;
|
|
2836
|
+
/**
|
|
2837
|
+
* StepOutput kinds that synthesize a full-frame detection when they run at
|
|
2838
|
+
* the tree ROOT (see `synthesizeRootDetection`) — classifier/embedding steps
|
|
2839
|
+
* promoted to root by `pipelineRunner.runDetailSubtree` (task 6b), e.g.
|
|
2840
|
+
* `clip-embedding`, `animal-classifier`. Detector (`'detections'`) roots take
|
|
2841
|
+
* the normal path; `'text'`/`'mask'` roots are out of scope and still
|
|
2842
|
+
* `continue` with no output, matching pre-existing behaviour.
|
|
2843
|
+
*/
|
|
2844
|
+
function isEnrichmentOutput(output) {
|
|
2845
|
+
return output.kind === "classifications" || output.kind === "embedding";
|
|
2846
|
+
}
|
|
4144
2847
|
var PipelineExecutor = class {
|
|
4145
2848
|
opts;
|
|
4146
2849
|
constructor(opts) {
|
|
@@ -4158,6 +2861,23 @@ var PipelineExecutor = class {
|
|
|
4158
2861
|
logger.warn(message, { meta });
|
|
4159
2862
|
}
|
|
4160
2863
|
/**
|
|
2864
|
+
* Surface WHY a face embedding was skipped — a detected face simply too small
|
|
2865
|
+
* (distance / night) is the common silent cause of "faces stopped collecting".
|
|
2866
|
+
* Reports the detection-space short side AND, when the skip was decided on the
|
|
2867
|
+
* native crop, the NATIVE face size so an operator sees the size that actually
|
|
2868
|
+
* gated. Throttled so it isn't per-frame spam.
|
|
2869
|
+
*/
|
|
2870
|
+
warnFaceTooSmall(child, detShortSidePx, minFaceSizePx, nativeFaceShortSidePx) {
|
|
2871
|
+
this.throttledWarn(`face-small:${child.stepId}`, 3e4, "face detected but too small to embed — skipping", {
|
|
2872
|
+
step: child.stepId,
|
|
2873
|
+
model: child.modelId,
|
|
2874
|
+
faceShortSidePx: Math.round(detShortSidePx),
|
|
2875
|
+
minFaceSizePx,
|
|
2876
|
+
gatedOn: nativeFaceShortSidePx === void 0 ? "detection" : "native",
|
|
2877
|
+
...nativeFaceShortSidePx === void 0 ? {} : { nativeFaceShortSidePx: Math.round(nativeFaceShortSidePx) }
|
|
2878
|
+
});
|
|
2879
|
+
}
|
|
2880
|
+
/**
|
|
4161
2881
|
* Execute the full pipeline tree against a frame.
|
|
4162
2882
|
*
|
|
4163
2883
|
* `rootInput` carries the frame in its source representation
|
|
@@ -4170,7 +2890,7 @@ var PipelineExecutor = class {
|
|
|
4170
2890
|
*
|
|
4171
2891
|
* @returns FrameResult + optional trace
|
|
4172
2892
|
*/
|
|
4173
|
-
async run(tree, rootInput, fullFrameJpegProvider, imageWidth, imageHeight, deviceId, runOpts) {
|
|
2893
|
+
async run(tree, rootInput, fullFrameJpegProvider, imageWidth, imageHeight, deviceId, runOpts, nativeFaceCropProvider) {
|
|
4174
2894
|
const startMs = Date.now();
|
|
4175
2895
|
const verbosity = runOpts?.traceVerbosity ?? "off";
|
|
4176
2896
|
const traceBuilder = new ExecutionTraceBuilder(verbosity, deviceId, imageWidth, imageHeight, this.opts.engineRuntime);
|
|
@@ -4193,7 +2913,27 @@ var PipelineExecutor = class {
|
|
|
4193
2913
|
const rootStart = Date.now();
|
|
4194
2914
|
const rootOutput = await this.executeStep(rootStep, rootInput, imageWidth, imageHeight, "full-frame", void 0, void 0, traceBuilder, stepTimings, poolAgg);
|
|
4195
2915
|
const rootMs = Date.now() - rootStart;
|
|
4196
|
-
if (rootOutput.kind !== "detections")
|
|
2916
|
+
if (rootOutput.kind !== "detections") {
|
|
2917
|
+
if (isEnrichmentOutput(rootOutput)) {
|
|
2918
|
+
const mutable = this.synthesizeRootDetection(rootOutput, rootStep, idGen, rootMs, imageWidth, imageHeight);
|
|
2919
|
+
applyChildOutput(mutable, rootStep, rootOutput, rootMs, ctx);
|
|
2920
|
+
try {
|
|
2921
|
+
await this.executeChildren(rootStep.children, mutable, fullFrameJpegProvider, imageWidth, imageHeight, traceBuilder, stepTimings, ctx, poolAgg, runOpts?.plane, nativeFaceCropProvider);
|
|
2922
|
+
} catch (err) {
|
|
2923
|
+
this.opts.logger?.warn("Pipeline child execution failed — keeping parent detection", {
|
|
2924
|
+
tags: { deviceId },
|
|
2925
|
+
meta: {
|
|
2926
|
+
rootStepId: rootStep.stepId,
|
|
2927
|
+
parentClass: mutable.macroClass,
|
|
2928
|
+
parentScore: mutable.score,
|
|
2929
|
+
error: errMsg(err)
|
|
2930
|
+
}
|
|
2931
|
+
});
|
|
2932
|
+
}
|
|
2933
|
+
firstLevel.push(mutable);
|
|
2934
|
+
}
|
|
2935
|
+
continue;
|
|
2936
|
+
}
|
|
4197
2937
|
for (const det of rootOutput.detections) {
|
|
4198
2938
|
const mutable = toMutableRootDetection(det, rootStep, idGen, rootMs);
|
|
4199
2939
|
if (rootStep.definition.classMap) {
|
|
@@ -4208,7 +2948,7 @@ var PipelineExecutor = class {
|
|
|
4208
2948
|
continue;
|
|
4209
2949
|
}
|
|
4210
2950
|
try {
|
|
4211
|
-
await this.executeChildren(rootStep.children, mutable, fullFrameJpegProvider, imageWidth, imageHeight, traceBuilder, stepTimings, ctx, poolAgg);
|
|
2951
|
+
await this.executeChildren(rootStep.children, mutable, fullFrameJpegProvider, imageWidth, imageHeight, traceBuilder, stepTimings, ctx, poolAgg, runOpts?.plane, nativeFaceCropProvider);
|
|
4212
2952
|
} catch (err) {
|
|
4213
2953
|
this.opts.logger?.warn("Pipeline child execution failed — keeping parent detection", {
|
|
4214
2954
|
tags: { deviceId },
|
|
@@ -4244,6 +2984,29 @@ var PipelineExecutor = class {
|
|
|
4244
2984
|
trace: traceBuilder.build(firstLevel.length + details.length)
|
|
4245
2985
|
};
|
|
4246
2986
|
}
|
|
2987
|
+
/**
|
|
2988
|
+
* Build a full-frame `MutableObjectDetection` for a ROOT step whose engine
|
|
2989
|
+
* output is a classification/embedding rather than detections — there is
|
|
2990
|
+
* no detector to bound a smaller box, so the synthesized detection covers
|
|
2991
|
+
* the entire input frame (`bbox = [0, 0, imageWidth, imageHeight]`).
|
|
2992
|
+
* `macroClass` is the step's own output semantic (`outputClasses[0]`,
|
|
2993
|
+
* falling back to `inputClasses[0]` when `outputClasses` is empty, then
|
|
2994
|
+
* the step id as a last resort so the field is never empty). `score` is
|
|
2995
|
+
* the output's own confidence when it has one (the classifier's top-1
|
|
2996
|
+
* score) — an embedding carries no confidence, so it defaults to 1.
|
|
2997
|
+
*/
|
|
2998
|
+
synthesizeRootDetection(output, rootStep, idGen, rootStepLatencyMs, imageWidth, imageHeight) {
|
|
2999
|
+
return toMutableRootDetection({
|
|
3000
|
+
class: rootStep.definition.outputClasses[0] ?? rootStep.definition.inputClasses?.[0] ?? rootStep.stepId,
|
|
3001
|
+
score: output.kind === "classifications" ? output.classifications[0]?.score ?? 1 : 1,
|
|
3002
|
+
bbox: [
|
|
3003
|
+
0,
|
|
3004
|
+
0,
|
|
3005
|
+
imageWidth,
|
|
3006
|
+
imageHeight
|
|
3007
|
+
]
|
|
3008
|
+
}, rootStep, idGen, rootStepLatencyMs);
|
|
3009
|
+
}
|
|
4247
3010
|
async executeStep(step, input, inputWidth, inputHeight, inputType, _parentBbox, parentClass, traceBuilder, stepTimings, poolAgg) {
|
|
4248
3011
|
const preprocessStart = Date.now();
|
|
4249
3012
|
const preprocessMs = Date.now() - preprocessStart;
|
|
@@ -4310,8 +3073,9 @@ var PipelineExecutor = class {
|
|
|
4310
3073
|
});
|
|
4311
3074
|
return output;
|
|
4312
3075
|
}
|
|
4313
|
-
async executeChildren(children, parentDetection, fullFrameJpegProvider, imageWidth, imageHeight, traceBuilder, stepTimings, ctx, poolAgg) {
|
|
3076
|
+
async executeChildren(children, parentDetection, fullFrameJpegProvider, imageWidth, imageHeight, traceBuilder, stepTimings, ctx, poolAgg, plane, nativeFaceCropProvider) {
|
|
4314
3077
|
for (const child of children) {
|
|
3078
|
+
if (plane === "frame" && child.definition.inputClasses !== null) continue;
|
|
4315
3079
|
if (!this.matchesInputClasses(parentDetection.macroClass, child.inputClasses)) continue;
|
|
4316
3080
|
const minParentScore = child.settings?.minParentScore ?? child.definition?.defaultMinParentScore;
|
|
4317
3081
|
if (typeof minParentScore === "number" && parentDetection.score < minParentScore) continue;
|
|
@@ -4325,23 +3089,43 @@ var PipelineExecutor = class {
|
|
|
4325
3089
|
let cropH;
|
|
4326
3090
|
if (modelEntry?.faceAlignment) {
|
|
4327
3091
|
const minFaceSize = typeof child.settings?.["minFaceSize"] === "number" ? child.settings["minFaceSize"] : DEFAULT_MIN_FACE_SIZE_PX;
|
|
4328
|
-
const
|
|
4329
|
-
if (
|
|
4330
|
-
this.
|
|
4331
|
-
step: child.stepId,
|
|
4332
|
-
model: child.modelId,
|
|
4333
|
-
faceShortSidePx: Math.round(faceShortSide),
|
|
4334
|
-
minFaceSizePx: minFaceSize
|
|
4335
|
-
});
|
|
3092
|
+
const detShortSide = bboxShortSide(parentDetection.bbox);
|
|
3093
|
+
if (detShortSide < MIN_FACE_DETECTION_FLOOR_PX) {
|
|
3094
|
+
this.warnFaceTooSmall(child, detShortSide, minFaceSize, void 0);
|
|
4336
3095
|
continue;
|
|
4337
3096
|
}
|
|
4338
3097
|
const lms = parentDetection.landmarks;
|
|
4339
3098
|
if (lms && lms.length >= 5) {
|
|
4340
|
-
|
|
3099
|
+
if (!nativeFaceCropProvider && detShortSide < minFaceSize) {
|
|
3100
|
+
this.warnFaceTooSmall(child, detShortSide, minFaceSize, void 0);
|
|
3101
|
+
continue;
|
|
3102
|
+
}
|
|
3103
|
+
const aligned = await buildAlignedFaceCrop({
|
|
3104
|
+
fullFrameJpeg,
|
|
3105
|
+
faceBbox: parentDetection.bbox,
|
|
3106
|
+
landmarksImageSpace: lms,
|
|
3107
|
+
imageWidth,
|
|
3108
|
+
imageHeight,
|
|
3109
|
+
outSize: modelEntry.inputSize.width,
|
|
3110
|
+
...nativeFaceCropProvider ? {
|
|
3111
|
+
nativeCropProvider: nativeFaceCropProvider,
|
|
3112
|
+
nativeMaxWidth: NATIVE_FACE_CROP_MAX_WIDTH
|
|
3113
|
+
} : {}
|
|
3114
|
+
});
|
|
3115
|
+
const gateShortSide = aligned.nativeFaceShortSidePx ?? detShortSide;
|
|
3116
|
+
if (gateShortSide < minFaceSize) {
|
|
3117
|
+
this.warnFaceTooSmall(child, detShortSide, minFaceSize, aligned.source === "native" ? gateShortSide : void 0);
|
|
3118
|
+
continue;
|
|
3119
|
+
}
|
|
4341
3120
|
cropJpegBuf = aligned.jpeg;
|
|
4342
3121
|
cropW = aligned.width;
|
|
4343
3122
|
cropH = aligned.height;
|
|
3123
|
+
parentDetection.faceAlignedCrop = cropJpegBuf.toString("base64");
|
|
4344
3124
|
} else {
|
|
3125
|
+
if (detShortSide < minFaceSize) {
|
|
3126
|
+
this.warnFaceTooSmall(child, detShortSide, minFaceSize, void 0);
|
|
3127
|
+
continue;
|
|
3128
|
+
}
|
|
4345
3129
|
const crop = await cropJpeg(fullFrameJpeg, parentDetection.bbox, imageWidth, imageHeight);
|
|
4346
3130
|
cropJpegBuf = crop.jpeg;
|
|
4347
3131
|
cropW = crop.width;
|
|
@@ -4371,7 +3155,7 @@ var PipelineExecutor = class {
|
|
|
4371
3155
|
y: l.y + py1
|
|
4372
3156
|
}));
|
|
4373
3157
|
}
|
|
4374
|
-
await this.executeChildren(child.children, detail, fullFrameJpegProvider, imageWidth, imageHeight, traceBuilder, stepTimings, ctx, poolAgg);
|
|
3158
|
+
await this.executeChildren(child.children, detail, fullFrameJpegProvider, imageWidth, imageHeight, traceBuilder, stepTimings, ctx, poolAgg, plane, nativeFaceCropProvider);
|
|
4375
3159
|
}
|
|
4376
3160
|
}
|
|
4377
3161
|
} catch (err) {
|
|
@@ -5863,6 +4647,38 @@ var DetectionPipelineProvider = class DetectionPipelineProvider {
|
|
|
5863
4647
|
debug: { totalInferenceMs: 0 }
|
|
5864
4648
|
};
|
|
5865
4649
|
}
|
|
4650
|
+
/**
|
|
4651
|
+
* Build the design-B native-face-crop provider for a runtime frame. Present
|
|
4652
|
+
* ONLY when the frame arrived as a retention `frameHandle` (the key the owning
|
|
4653
|
+
* pipeline-runner registered the frame's native surface under) and the sibling
|
|
4654
|
+
* `pipeline-runner` cap is reachable via `ctx.api`. Each call is a best-effort
|
|
4655
|
+
* round-trip pinned to the frame's OWNING node (`nodePin(handle.nodeId)` — the
|
|
4656
|
+
* node that decoded it, i.e. THIS node for a local frame, never a cross-node
|
|
4657
|
+
* hop) that returns the native ROI or `null` on a miss; the executor then
|
|
4658
|
+
* falls back to the downscaled crop, so behaviour is never worse than before.
|
|
4659
|
+
*/
|
|
4660
|
+
buildNativeFaceCropProvider(handle) {
|
|
4661
|
+
if (!handle) return void 0;
|
|
4662
|
+
const api = this.addonCtx?.api;
|
|
4663
|
+
if (!api) return void 0;
|
|
4664
|
+
return async (roi, maxWidth) => {
|
|
4665
|
+
try {
|
|
4666
|
+
const native = await api.pipelineRunner.getNativeCrop.query({
|
|
4667
|
+
handle,
|
|
4668
|
+
bbox: roi,
|
|
4669
|
+
...maxWidth !== void 0 ? { maxWidth } : {}
|
|
4670
|
+
}, nodePin(handle.nodeId));
|
|
4671
|
+
if (!native || native.width <= 0 || native.height <= 0) return null;
|
|
4672
|
+
return {
|
|
4673
|
+
bytes: native.bytes,
|
|
4674
|
+
width: native.width,
|
|
4675
|
+
height: native.height
|
|
4676
|
+
};
|
|
4677
|
+
} catch {
|
|
4678
|
+
return null;
|
|
4679
|
+
}
|
|
4680
|
+
};
|
|
4681
|
+
}
|
|
5866
4682
|
async runPipeline(input, onProgress) {
|
|
5867
4683
|
const nodeId = this.addonCtx?.kernel?.localNodeId ?? "hub";
|
|
5868
4684
|
const sessionId = input.sessionId ?? `run-${Date.now()}-${Math.random().toString(36).slice(2, 10)}`;
|
|
@@ -6119,7 +4935,11 @@ var DetectionPipelineProvider = class DetectionPipelineProvider {
|
|
|
6119
4935
|
const effectiveDeviceId = input.deviceId ?? 0;
|
|
6120
4936
|
const deviceOverrides = effectiveDeviceId > 0 ? await this.readDeviceStore(effectiveDeviceId) : {};
|
|
6121
4937
|
const effectiveTree = Object.keys(deviceOverrides).length > 0 ? applyDeviceOverridesToTree(tree, "object-detection", deviceOverrides) : tree;
|
|
6122
|
-
const
|
|
4938
|
+
const nativeFaceCropProvider = this.buildNativeFaceCropProvider(input.frameHandle);
|
|
4939
|
+
const { result, trace } = await executor.run(effectiveTree, rootInput, jpegProvider, imageWidth, imageHeight, effectiveDeviceId, {
|
|
4940
|
+
traceVerbosity: isRuntime ? this.eventBus ? "summary" : "off" : "full",
|
|
4941
|
+
plane: input.plane
|
|
4942
|
+
}, nativeFaceCropProvider);
|
|
6123
4943
|
if (isRuntime) {
|
|
6124
4944
|
if (trace && this.eventBus) this.eventBus.emit(createEvent(EventCategory.PipelineTrace, {
|
|
6125
4945
|
type: "device",
|