EXCALORIO commited on
Commit
63bec12
·
verified ·
1 Parent(s): fd41906

ficha: que se extrae, con que precision, y los videos

Browse files
Files changed (1) hide show
  1. README.md +14 -26
README.md CHANGED
@@ -23,7 +23,7 @@ yet: our detector scores **69.7 %** against **73.8 % for never answering at all*
23
  everything that *is* extracted along the way.
24
 
25
  Everything below is **generated by us** from the clips: skeletons, masks, normals, alphas, threat
26
- cones, piste maps. Every video and every figure on this page comes out of the pipeline - 22
27
  videos in all - and each one carries **its measured error printed on the frame**, not just its name.
28
 
29
  ---
@@ -295,18 +295,6 @@ The same thing zoomed into the two panels that carry the measurement, because at
295
 
296
  ***The piste from above, to scale**, with the regulation lines. Each fencer is a thick dot, the trail behind is where they have been, and the events - march, retreat, lunge, halt - appear labelled as they happen.*
297
 
298
- <video src="https://huggingface.co/datasets/EXCALORIO/fencingly-datos/resolve/main/media/trace_ForcefulAntiqueAlbacoretuna_public.mp4" controls loop muted playsinline width="100%"></video>
299
-
300
- [trace_ForcefulAntiqueAlbacoretuna_public.mp4 - MP4, 7.0 MB](media/trace_ForcefulAntiqueAlbacoretuna_public.mp4)
301
-
302
-
303
- <video src="https://huggingface.co/datasets/EXCALORIO/fencingly-datos/resolve/main/media/trace_UglyNiceAplomadofalcon_public.mp4" controls loop muted playsinline width="100%"></video>
304
-
305
- [trace_UglyNiceAplomadofalcon_public.mp4 - MP4, 4.1 MB](media/trace_UglyNiceAplomadofalcon_public.mp4)
306
-
307
-
308
- At quarter speed, for judging the detection lag frame by frame: [ForcefulAntiqueAlbacoretuna](media/trace_ForcefulAntiqueAlbacoretuna_slow_public.mp4) | [UglyNiceAplomadofalcon](media/trace_UglyNiceAplomadofalcon_slow_public.mp4)
309
-
310
  **What is measured now**, each with its own denominator, never pooled:
311
 
312
  | | value | out of |
@@ -410,8 +398,8 @@ against, so there is nothing to ship. `halt` is **not** a parry: it means the fe
410
  | | |
411
  |---|---:|
412
  | clips in the corpus | 8,559 |
413
- | clips with 308-point poses today | **38** |
414
- | movement events extracted | **650** over 22 clips |
415
  | frames per clip, median | 307 |
416
  | resolution | 81 % at 1280x720, 18 % at 1920x1080, 0.8 % at 854x480 |
417
  | cost, measured over 50 clips | **17.3 s/clip** to detect and pair (83 % of it the detector), plus **28 s/clip** for the four dense layers |
@@ -427,17 +415,17 @@ against, so there is nothing to ship. `halt` is **not** a parry: it means the fe
427
 
428
  | field | shape on disk | what it is | format | measured precision |
429
  |---|---|---|---|---|
430
- | `KP` | `(316, 2, 308, 2)` | the 308 body keypoints of both fencers, in pixels | px, float32 | **2.26 px** torso error at 384x288 instead of 1024 (7.8x cheaper). **96.0 %** of delivered poses are MEASURED, not filled in |
431
- | `SC` | `(316, 2, 308)` | per-keypoint confidence | 0-1, float32 | used with a 0.3 threshold. Points that fell outside their own person's box are zeroed: there were **10,140** of them and they sat on the other fencer |
432
- | `origen` | `(316, 2)` | how the pose was obtained: 0 measured, 1 interpolated, 2 held, -1 none | code, int8 | **0 held poses** in what is delivered. Without this field the data cannot be weighted, and it **cannot be recovered later** |
433
- | `hueco` | `(316, 2)` | length in frames of the gap an invented pose comes from | frames, int16 | **this row used to quote 2.2 % of torso length for any gap <= 24 frames, and that was wrong**: 2.2 % is the cost of a **1-frame** gap. Re-measured over the gaps that actually get filled, the whole <= 24 band costs **6.33 %** (n = 271), and a 24-frame gap costs **25.7 %**. Beyond 24 frames **there is no measurement at all** |
434
- | `B` | `(316, 2, 4)` | each fencer's box, normalised 0-1, xyxy | fraction of frame, float32 | no delivered frame jumps more than **0.60 box widths** (worst measured 0.563). No box is a border sliver, and none comes from a track that was never seen whole |
435
- | `Be` | `(316, 2, 4)` | ESTIMATED box: where the fencer should be in a frame the detector missed | fraction of frame, float32 | used only to aim a crop re-detection; its own error is not measured separately |
436
- | `pres` | `(316, 2)` | whether that fencer's pose in that frame is MEASURED | bool | **0** delivered frames without both fencers, across 28 clips |
437
- | `mask` | `(316,)` | usable frames: close-ups, replays and cross-fades removed | bool | a cross-fade is told from a zoom by the residual relative to contrast: **0.219 vs 0.062** |
438
  | `filtro_vacias` | `()` | measured poses dropped for coming out empty | count, int32 | a pose with fewer than **20 useful body points out of 63** cannot serve as an anchor |
439
- | `filtro_fuera_de_caja` | `(), in 21 of 38 files` | keypoints zeroed for falling outside their own box | count, int32 | **10,140** across 28 clips |
440
- | `recortados_por_pose` | `(), in 21 of 38 files` | frames trimmed off the ends for not having both poses | count, int32 | **80** across 22 clips, which is what leaves **0 held poses** |
441
  | `W` | `()` | video width | px, int | **81 %** of the corpus is 1280x720 |
442
  | `H` | `()` | video height | px, int | 18 % is 1920x1080, 0.8 % is 854x480 |
443
  | `T` | `()` | frames in the clip | frames, int | median **307**. The previous pipeline stored a **129-frame window**, not the whole clip |
@@ -455,7 +443,7 @@ against, so there is nothing to ship. `halt` is **not** a parry: it means the fe
455
 
456
  ### 3 - The movement events: one row per event, not per frame
457
 
458
- **650 events over 22 clips** so far. Not published here as a file - the schema below is the contract, and the rows ship in batches of 500.
459
 
460
  | column | what it is | format | measured precision |
461
  |---|---|---|---|
 
23
  everything that *is* extracted along the way.
24
 
25
  Everything below is **generated by us** from the clips: skeletons, masks, normals, alphas, threat
26
+ cones, piste maps. Every video and every figure on this page comes out of the pipeline - 38
27
  videos in all - and each one carries **its measured error printed on the frame**, not just its name.
28
 
29
  ---
 
295
 
296
  ***The piste from above, to scale**, with the regulation lines. Each fencer is a thick dot, the trail behind is where they have been, and the events - march, retreat, lunge, halt - appear labelled as they happen.*
297
 
 
 
 
 
 
 
 
 
 
 
 
 
298
  **What is measured now**, each with its own denominator, never pooled:
299
 
300
  | | value | out of |
 
398
  | | |
399
  |---|---:|
400
  | clips in the corpus | 8,559 |
401
+ | clips with 308-point poses today | **661** |
402
+ | movement events extracted | **1,717** over 65 clips |
403
  | frames per clip, median | 307 |
404
  | resolution | 81 % at 1280x720, 18 % at 1920x1080, 0.8 % at 854x480 |
405
  | cost, measured over 50 clips | **17.3 s/clip** to detect and pair (83 % of it the detector), plus **28 s/clip** for the four dense layers |
 
415
 
416
  | field | shape on disk | what it is | format | measured precision |
417
  |---|---|---|---|---|
418
+ | `KP` | `(300, 2, 308, 2)` | the 308 body keypoints of both fencers, in pixels | px, float32 | **2.26 px** torso error at 384x288 instead of 1024 (7.8x cheaper). **96.0 %** of delivered poses are MEASURED, not filled in |
419
+ | `SC` | `(300, 2, 308)` | per-keypoint confidence | 0-1, float32 | used with a 0.3 threshold. Points that fell outside their own person's box are zeroed: there were **10,140** of them and they sat on the other fencer |
420
+ | `origen` | `(300, 2)` | how the pose was obtained: 0 measured, 1 interpolated, 2 held, -1 none | code, int8 | **0 held poses** in what is delivered. Without this field the data cannot be weighted, and it **cannot be recovered later** |
421
+ | `hueco` | `(300, 2)` | length in frames of the gap an invented pose comes from | frames, int16 | **this row used to quote 2.2 % of torso length for any gap <= 24 frames, and that was wrong**: 2.2 % is the cost of a **1-frame** gap. Re-measured over the gaps that actually get filled, the whole <= 24 band costs **6.33 %** (n = 271), and a 24-frame gap costs **25.7 %**. Beyond 24 frames **there is no measurement at all** |
422
+ | `B` | `(300, 2, 4)` | each fencer's box, normalised 0-1, xyxy | fraction of frame, float32 | no delivered frame jumps more than **0.60 box widths** (worst measured 0.563). No box is a border sliver, and none comes from a track that was never seen whole |
423
+ | `Be` | `(300, 2, 4)` | ESTIMATED box: where the fencer should be in a frame the detector missed | fraction of frame, float32 | used only to aim a crop re-detection; its own error is not measured separately |
424
+ | `pres` | `(300, 2)` | whether that fencer's pose in that frame is MEASURED | bool | **0** delivered frames without both fencers, across 28 clips |
425
+ | `mask` | `(300,)` | usable frames: close-ups, replays and cross-fades removed | bool | a cross-fade is told from a zoom by the residual relative to contrast: **0.219 vs 0.062** |
426
  | `filtro_vacias` | `()` | measured poses dropped for coming out empty | count, int32 | a pose with fewer than **20 useful body points out of 63** cannot serve as an anchor |
427
+ | `filtro_fuera_de_caja` | `(), in 21 of 661 files` | keypoints zeroed for falling outside their own box | count, int32 | **10,140** across 28 clips |
428
+ | `recortados_por_pose` | `(), in 21 of 661 files` | frames trimmed off the ends for not having both poses | count, int32 | **80** across 22 clips, which is what leaves **0 held poses** |
429
  | `W` | `()` | video width | px, int | **81 %** of the corpus is 1280x720 |
430
  | `H` | `()` | video height | px, int | 18 % is 1920x1080, 0.8 % is 854x480 |
431
  | `T` | `()` | frames in the clip | frames, int | median **307**. The previous pipeline stored a **129-frame window**, not the whole clip |
 
443
 
444
  ### 3 - The movement events: one row per event, not per frame
445
 
446
+ **1717 events over 65 clips** so far. Not published here as a file - the schema below is the contract, and the rows ship in batches of 500.
447
 
448
  | column | what it is | format | measured precision |
449
  |---|---|---|---|