ficha: que se extrae, con que precision, y los videos
Browse files
README.md
CHANGED
|
@@ -18,10 +18,12 @@ viewer: false
|
|
| 18 |
We are trying to learn **who has right of way** in a foil action. Foil is a weapon of *convention*:
|
| 19 |
the valid touch is not simply the first one to land, so the answer **cannot be read off the scoring
|
| 20 |
lights**. That is the whole reason this project exists - and it is the one thing that does not work
|
| 21 |
-
yet
|
|
|
|
|
|
|
| 22 |
|
| 23 |
Everything below is **generated by us** from the clips: skeletons, masks, normals, alphas, threat
|
| 24 |
-
cones, piste maps. Every video and every figure on this page comes out of the pipeline -
|
| 25 |
videos in all - and each one carries **its measured error printed on the frame**, not just its name.
|
| 26 |
|
| 27 |
---
|
|
@@ -52,7 +54,7 @@ Two fencers, 308 points each, every frame. Green is side 0 (image left), orange
|
|
| 52 |
|
| 53 |
<video src="https://huggingface.co/datasets/EXCALORIO/fencingly-datos/resolve/main/media/pose_proudDungBeetleBoglenTerrier.mp4" controls loop muted playsinline width="100%"></video>
|
| 54 |
|
| 55 |
-
[pose_proudDungBeetleBoglenTerrier.mp4 - MP4, 3.
|
| 56 |
|
| 57 |
More clips, same layer: [uptightSlothEveningBat](media/pose_uptightSlothEveningBat.mp4)
|
| 58 |
|
|
@@ -186,11 +188,11 @@ The yellow cross is where **the two cones cross each other**. That is the test -
|
|
| 186 |
|
| 187 |
<video src="https://huggingface.co/datasets/EXCALORIO/fencingly-datos/resolve/main/media/foil_BabyishIdioticCockerspaniel.mp4" controls loop muted playsinline width="100%"></video>
|
| 188 |
|
| 189 |
-
[foil_BabyishIdioticCockerspaniel.mp4 - MP4, 2.
|
| 190 |
|
| 191 |
<video src="https://huggingface.co/datasets/EXCALORIO/fencingly-datos/resolve/main/media/foil_proudDungBeetleBoglenTerrier.mp4" controls loop muted playsinline width="100%"></video>
|
| 192 |
|
| 193 |
-
[foil_proudDungBeetleBoglenTerrier.mp4 - MP4, 3.
|
| 194 |
|
| 195 |
More clips, same layer: [uptightSlothEveningBat](media/foil_uptightSlothEveningBat.mp4)
|
| 196 |
|
|
@@ -206,8 +208,9 @@ More stills, same layer: [BabyishIdioticCockerspaniel_c0302](media/foil_BabyishI
|
|
| 206 |
|
| 207 |
- reach **1.10 m** wrist-to-tip. A foil blade is **0.90 m** on its own, and the first version used exactly that - which is wrong for a *threat* radius, because the threat starts at the wrist and the guard and the hand add the rest. It looked too short and it was
|
| 208 |
- the axis is the **forearm**, `elbow -> hand root`. Using the hand axis instead put the cone more than 40 deg off in **63.6 %** of frames; the forearm brings that to **15.7 %**
|
| 209 |
-
- the arcs are **measured, not chosen**: solid = p50 core **+-14.4 deg half-angle (28.8 deg full)**, dashed = p90 envelope **+-33.0 deg half-angle (66.0 deg full)**, both from |detected blade - forearm axis| over 31 blades. Every angle here is written as *half and full*, on purpose: that undeclared convention already cost this project five contradictory figures in one day
|
| 210 |
-
- when a blade is actually **detected** the field is centred on the
|
|
|
|
| 211 |
- **honest limit, broken down instead of pooled**: a blade line is found in **42.6 %** of crops at 1920x1080 and in **1.6 %** at 1280x720 - and **81 %** of the corpus is 720p. So most of the time the cone is a **geometric prior, not a detection**, and a single pooled percentage would hide which clips it is true for
|
| 212 |
- the thumb axis was tried as the blade direction and **failed**: lever arm of 2.4-4.8 px, jumps over 40 deg in 10-26 % of frames, against 0.9-2.1 deg for the forearm
|
| 213 |
|
|
@@ -269,70 +272,17 @@ March, retreat, lunge, halt, arm extension - one row per event, not per frame. T
|
|
| 269 |
|
| 270 |
### Which hand holds the weapon
|
| 271 |
|
| 272 |
-
|
| 273 |
|
| 274 |

|
| 275 |
|
| 276 |
-
- **1.000 accuracy over 19 of 19 fencers**
|
| 277 |
-
- the
|
|
|
|
| 278 |
- the glove was tried and **failed**: the whiteness feature reads the **sleeve** (pearson 0.681), and dark gloves exist. Paired separation **-0.002**, i.e. backwards
|
| 279 |
- what cannot be decided is flagged `uncertain` rather than guessed
|
| 280 |
|
| 281 |
|
| 282 |
-
### Right of way - the goal of the whole project, and it does not work yet
|
| 283 |
-
|
| 284 |
-
Foil is a weapon of **convention**: the valid touch is not simply the first one to land. When both lights come on, a referee decides who had **right of way** (FIE t.82-t.89). That is the label we are trying to learn, and this is the honest state of it.
|
| 285 |
-
|
| 286 |
-

|
| 287 |
-
|
| 288 |
-

|
| 289 |
-
|
| 290 |
-

|
| 291 |
-
|
| 292 |
-

|
| 293 |
-
|
| 294 |
-
<video src="https://huggingface.co/datasets/EXCALORIO/fencingly-datos/resolve/main/media/prio_video_right_of_way_KindEnchantingKid_c205_c262.mp4" controls loop muted playsinline width="100%"></video>
|
| 295 |
-
|
| 296 |
-
[prio_video_right_of_way_KindEnchantingKid_c205_c262.mp4 - MP4, 3.7 MB](media/prio_video_right_of_way_KindEnchantingKid_c205_c262.mp4)
|
| 297 |
-
|
| 298 |
-
|
| 299 |
-
**Who holds right of way in this frame** - our detector, against the baseline that costs nothing:
|
| 300 |
-
|
| 301 |
-
| | value | measured over |
|
| 302 |
-
|---|---:|---|
|
| 303 |
-
| our right-of-way detector | **69.7 %** | 462 observations |
|
| 304 |
-
| **never answering at all** - the free baseline it has to beat | **73.8 %** | the same 462 |
|
| 305 |
-
| the same detector on held-out control clips | **9.1 %** | 44 observations |
|
| 306 |
-
|
| 307 |
-
It ships **off**, with `PRIORIDAD = False` and the number next to the switch. **A detector that is beaten by declining to answer is not a detector.**
|
| 308 |
-
|
| 309 |
-
The free baseline is 73.8 % for a plain reason: **12 of the 21 annotated windows are *nobody* from start to finish**. Answering nothing scores that without knowing anything, and anything that does fire pays for it there.
|
| 310 |
-
|
| 311 |
-
Split in two, the failure has a shape - and it is the shape that says **no threshold sweep will fix this**:
|
| 312 |
-
|
| 313 |
-
| | detector | the crude box-motion baseline | answering nothing |
|
| 314 |
-
|---|---:|---:|---:|
|
| 315 |
-
| frames that **do** have an owner (n = 121) | **25.6 %** | 41.3 % | 0 % |
|
| 316 |
-
| frames with **no** owner (n = 341) | **85.3 %** | 32.3 % | **100 %** |
|
| 317 |
-
|
| 318 |
-
It is beaten on **both halves of the job** - worse than crude box motion when it has to name someone, worse than silence when there is nobody to name. There is no operating point where it wins.
|
| 319 |
-
|
| 320 |
-
The one encouraging number, stated precisely: when the detector **does** commit and the annotation also names an owner, **the side it picks is right 86.1 %** of the time (n = 36). But it only commits on **36 of those 121 frames**. *The side is good; the coverage is bad.* And on the held-out clips it does not answer at all: **0.0 %** of the held-out frames that have an owner get one.
|
| 321 |
-
|
| 322 |
-
Two caveats on those figures, because they are the kind that get quietly rounded away. The 462 are **observations, not distinct frames**: two annotation windows overlap on 7 frames, so there are 455 distinct ones, and both copies carry identical truth and output so nothing is biased - but 462 is not 455. And a re-run today gave **70.35 %** on the same 462, with 22 decisions differing and two clips moving in opposite directions. The verdict does not move: **both numbers lose to 73.8 %**.
|
| 323 |
-
|
| 324 |
-
### Why the scoring lights are not the answer either
|
| 325 |
-
|
| 326 |
-
| | value | measured over | unit |
|
| 327 |
-
|---|---:|---|---|
|
| 328 |
-
| *whoever lights up second wins* | 64.96 % | 1,815 clips | **who won the touch** |
|
| 329 |
-
| in one-light clips, *whoever lights up wins* | 96.19 % | one-light clips | **who won the touch** |
|
| 330 |
-
|
| 331 |
-
**These are a different question on a different unit and they do not compare to the 69.7 % above.** That row answers *who won the touch*, per clip; the detector answers *who holds right of way*, per frame. Putting them in one table - which an earlier version of this page did - invents a comparison that was never measured.
|
| 332 |
-
|
| 333 |
-
What they do show is that the lights are a **shortcut, not a solution**: in the clips where *second light wins* is false, a model leaning on it falls **below chance**, and one-light clips teach nothing about right of way at all. The audio beep was tried as a light detector and **failed**: **50.6 %** of the corpus has no audio track, and where it does, several pistes beep in the same hall.
|
| 334 |
-
|
| 335 |
-
|
| 336 |
### The open defect: sometimes one of the two fencers is the referee
|
| 337 |
|
| 338 |
In about **6 % of clips the pipeline delivers the REFEREE as one of the two fencers.** Four detection routes were tried and **all four failed, each with a number**:
|
|
@@ -361,7 +311,7 @@ promise, and a capability without its failures is a brochure.
|
|
| 361 |
| provenance of every point | **works** - and it is the part that cannot be recovered later |
|
| 362 |
| the five model channels | **works** - referee leakage 0.0000 |
|
| 363 |
| movement events | **works** - 90 % recall, `march -> lunge` confusion = 0 |
|
| 364 |
-
| weapon hand | **works on a held-out set** - 1.000 over 19
|
| 365 |
| piste rectified from its two edges | **works** - 20 of 22 clips |
|
| 366 |
| **origin** on the piste, measured not assumed | **works where the colour frontiers are visible** - 13 of 22 clips, held-out landmark to **p50 5 cm** |
|
| 367 |
| scoring lights | **works** - 90.5 % with a per-venue region of interest |
|
|
@@ -423,10 +373,11 @@ against, so there is nothing to ship. `halt` is **not** a parry: it means the fe
|
|
| 423 |
| channel | what it is | format | why it is there, measured |
|
| 424 |
|---|---|---|---|
|
| 425 |
| `pose` | 308 keypoints per fencer | (T, 2, 308, 2) float32 + (T, 2, 308) confidence | the index map is MEASURED: 5-6 shoulders, 7-8 elbows, 9-10 hips, 11-12 knees, 13-14 ankles, 15-20 feet, 21-41 hand A, 42-62 hand B, 63+ face. **245 of the 308 are face** and the project discards them, because fencers wear a mask |
|
| 426 |
-
| `seg` | body-part segmentation of each fencer | crop-sized uint8 class map | referee leakage into the mask measured **0.0000** |
|
| 427 |
-
| `normal` | surface normals | crop-sized float16, 3 channels, NaN outside the mask | the dense layer that **does** hold up for local geometry |
|
| 428 |
-
| `matting` | alpha, the soft silhouette | crop-sized float16, 0 outside the mask | at very low alpha the **edge of the blade** appears, which is what the foil detector uses |
|
| 429 |
-
| `pointmap` | the model's 3D point map, z channel | crop-sized float16, NaN outside | **kept but not used**: the
|
|
|
|
| 430 |
|
| 431 |
### 3 - The movement events: one row per event, not per frame
|
| 432 |
|
|
@@ -446,8 +397,8 @@ against, so there is nothing to ship. `halt` is **not** a parry: it means the fe
|
|
| 446 |
| `worst_pose_provenance` | the worst pose provenance any frame of the event rests on | code, int | an event resting on an interpolated pose is weaker, and here you can see it |
|
| 447 |
| `gap_beyond_measured` | whether the event uses a pose from a gap longer than 24 frames | bool | flags data whose error is **not** verifiable |
|
| 448 |
| `side_valid` | whether that side really is a fencer | bool | in **6.1 %** of clips the delivered side is the **REFEREE**, and **all four** detection routes tried have failed. This flag is a **lower bound** |
|
| 449 |
-
| `weapon_hand` | which hand holds the weapon: A (root 21) or B (root 42) | category | **1.000
|
| 450 |
-
| `campo_mu_gr / campo_sigma_gr / campo_origen` | the blade probability field: centre direction, width, and where it came from | deg, deg, code | `origen 0` = the blade was **SEEN**, so the centre is the **measured** blade angle and the
|
| 451 |
|
| 452 |
|
| 453 |
---
|
|
@@ -485,9 +436,12 @@ Where a number appears twice on this page with two different denominators, that
|
|
| 485 |
controls count **clips**, some count **fencers** and some count **frames**, and pooling them would
|
| 486 |
invent a figure that was never measured.
|
| 487 |
|
| 488 |
-
|
| 489 |
-
sign on a bigger set, a blade dispersion went from 14.0 deg to
|
| 490 |
-
|
|
|
|
|
|
|
|
|
|
| 491 |
|
| 492 |
---
|
| 493 |
|
|
|
|
| 18 |
We are trying to learn **who has right of way** in a foil action. Foil is a weapon of *convention*:
|
| 19 |
the valid touch is not simply the first one to land, so the answer **cannot be read off the scoring
|
| 20 |
lights**. That is the whole reason this project exists - and it is the one thing that does not work
|
| 21 |
+
yet: our detector scores **69.7 %** against **73.8 % for never answering at all**, so it ships
|
| 22 |
+
**off**. It is in the capability table below with that number next to it, and this page is about
|
| 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 |
---
|
|
|
|
| 54 |
|
| 55 |
<video src="https://huggingface.co/datasets/EXCALORIO/fencingly-datos/resolve/main/media/pose_proudDungBeetleBoglenTerrier.mp4" controls loop muted playsinline width="100%"></video>
|
| 56 |
|
| 57 |
+
[pose_proudDungBeetleBoglenTerrier.mp4 - MP4, 3.6 MB](media/pose_proudDungBeetleBoglenTerrier.mp4)
|
| 58 |
|
| 59 |
More clips, same layer: [uptightSlothEveningBat](media/pose_uptightSlothEveningBat.mp4)
|
| 60 |
|
|
|
|
| 188 |
|
| 189 |
<video src="https://huggingface.co/datasets/EXCALORIO/fencingly-datos/resolve/main/media/foil_BabyishIdioticCockerspaniel.mp4" controls loop muted playsinline width="100%"></video>
|
| 190 |
|
| 191 |
+
[foil_BabyishIdioticCockerspaniel.mp4 - MP4, 2.7 MB](media/foil_BabyishIdioticCockerspaniel.mp4)
|
| 192 |
|
| 193 |
<video src="https://huggingface.co/datasets/EXCALORIO/fencingly-datos/resolve/main/media/foil_proudDungBeetleBoglenTerrier.mp4" controls loop muted playsinline width="100%"></video>
|
| 194 |
|
| 195 |
+
[foil_proudDungBeetleBoglenTerrier.mp4 - MP4, 3.7 MB](media/foil_proudDungBeetleBoglenTerrier.mp4)
|
| 196 |
|
| 197 |
More clips, same layer: [uptightSlothEveningBat](media/foil_uptightSlothEveningBat.mp4)
|
| 198 |
|
|
|
|
| 208 |
|
| 209 |
- reach **1.10 m** wrist-to-tip. A foil blade is **0.90 m** on its own, and the first version used exactly that - which is wrong for a *threat* radius, because the threat starts at the wrist and the guard and the hand add the rest. It looked too short and it was
|
| 210 |
- the axis is the **forearm**, `elbow -> hand root`. Using the hand axis instead put the cone more than 40 deg off in **63.6 %** of frames; the forearm brings that to **15.7 %**
|
| 211 |
+
- the arcs are **measured, not chosen**: solid = p50 core **+-14.4 deg half-angle (28.8 deg full)**, dashed = p90 envelope **+-33.0 deg half-angle (66.0 deg full)**, both from |detected blade - forearm axis| over 31 blades. Every angle here is written as *half and full*, on purpose: that undeclared convention already cost this project five contradictory figures in one day. Those two arcs are the pair for an **assumed** direction; where the blade was measured there is only one arc, at +-8.6 deg, because there is no measured p50 for that provenance and inventing one would be a third figure
|
| 212 |
+
- when a blade is actually **detected** the field is **centred on the measured blade itself** (`origen 0`) and narrows to **+-8.6 deg**, which is **3.8x** narrower than the +-33.0 deg it uses when the direction is only assumed. Centre and width are **one decision**, not two: the 33.0 is by definition the p90 of |blade - forearm|, so the moment the centre moves onto the blade that percentile no longer applies, and 8.6 is the p90 of what is left over
|
| 213 |
+
- **two figures on this page were wrong and are retired with their measurements**, because that is the only way a correction is worth anything. The **layer videos used to centre the cone on the forearm while the caption said it was centred on the measured blade** - in the frame a reader pointed at, the blade sat **21.4 deg** off the axis, outside the solid p50 core, and it was not a rare frame: **26 of 60**. And the width for a seen blade was published as **4.2 deg**, which is the frame-to-frame tremor of the blade's *direction* (70 pairs over 12 bouts, itself a correction of an earlier 14.0 from 8 pairs of one bout) - a real number, but **not** the width of a wedge whose apex is the hand: the detected blade line misses the hand root by 7 px median at only 0.24 of the reach, and a few pixels that close to the apex subtend many degrees. At 4.2 deg the cone **did not contain the blade it was centred on** in 21 of 60 frames. `temblor_hoja_gr` still ships, as its own field, because it measures something real - just not this
|
| 214 |
- **honest limit, broken down instead of pooled**: a blade line is found in **42.6 %** of crops at 1920x1080 and in **1.6 %** at 1280x720 - and **81 %** of the corpus is 720p. So most of the time the cone is a **geometric prior, not a detection**, and a single pooled percentage would hide which clips it is true for
|
| 215 |
- the thumb axis was tried as the blade direction and **failed**: lever arm of 2.4-4.8 px, jumps over 40 deg in 10-26 % of frames, against 0.9-2.1 deg for the forearm
|
| 216 |
|
|
|
|
| 272 |
|
| 273 |
### Which hand holds the weapon
|
| 274 |
|
| 275 |
+
What decides is a **cascade**, and its first rung is the rule that carries it: **in the on-guard stance the front foot is on the same side as the weapon arm** (FIE t.115.1). It is counter-intuitive and it is what works. If the front-foot vote is not firm, an on-guard-and-in-action gate takes over; what neither settles is flagged `uncertain`.
|
| 276 |
|
| 277 |

|
| 278 |
|
| 279 |
+
- **the cascade: 1.000 accuracy over 19 of 19 fencers** of a held-out set that played no part in choosing the rule, and **0.949 over all 59** fencers with ground truth
|
| 280 |
+
- **the front-foot rule on its own**, with a firm vote: **12 of 12 with no failure, covering 12 of those 19**. Ungated it scores **0.895 over 19** and 0.881 over 59. The 1.000 belongs to the cascade and the coverage to the rule, and they are not the same claim: an earlier version of this page merged them into one sentence
|
| 281 |
+
- the previous rule, *the hand closest to the opponent*, scored **0.831 over 59** - wrong about one time in six, and 0.741 on the first 27 annotated
|
| 282 |
- the glove was tried and **failed**: the whiteness feature reads the **sleeve** (pearson 0.681), and dark gloves exist. Paired separation **-0.002**, i.e. backwards
|
| 283 |
- what cannot be decided is flagged `uncertain` rather than guessed
|
| 284 |
|
| 285 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 286 |
### The open defect: sometimes one of the two fencers is the referee
|
| 287 |
|
| 288 |
In about **6 % of clips the pipeline delivers the REFEREE as one of the two fencers.** Four detection routes were tried and **all four failed, each with a number**:
|
|
|
|
| 311 |
| provenance of every point | **works** - and it is the part that cannot be recovered later |
|
| 312 |
| the five model channels | **works** - referee leakage 0.0000 |
|
| 313 |
| movement events | **works** - 90 % recall, `march -> lunge` confusion = 0 |
|
| 314 |
+
| weapon hand | **works on a held-out set** - the cascade 1.000 over 19 of 19; its front-foot rung alone covers 12 of those 19 |
|
| 315 |
| piste rectified from its two edges | **works** - 20 of 22 clips |
|
| 316 |
| **origin** on the piste, measured not assumed | **works where the colour frontiers are visible** - 13 of 22 clips, held-out landmark to **p50 5 cm** |
|
| 317 |
| scoring lights | **works** - 90.5 % with a per-venue region of interest |
|
|
|
|
| 373 |
| channel | what it is | format | why it is there, measured |
|
| 374 |
|---|---|---|---|
|
| 375 |
| `pose` | 308 keypoints per fencer | (T, 2, 308, 2) float32 + (T, 2, 308) confidence | the index map is MEASURED: 5-6 shoulders, 7-8 elbows, 9-10 hips, 11-12 knees, 13-14 ankles, 15-20 feet, 21-41 hand A, 42-62 hand B, 63+ face. **245 of the 308 are face** and the project discards them, because fencers wear a mask |
|
| 376 |
+
| `seg` | body-part segmentation of each fencer | crop-sized uint8 class map | referee leakage into the mask measured **0.0000**. In the layer videos the channel is blended **inside the mask only** and the frame outside it is merely darkened, so you can check by eye that the mask landed on the right person and that nothing around it moved |
|
| 377 |
+
| `normal` | surface normals | crop-sized float16, 3 channels, NaN outside the mask | one axis per colour channel; the dense layer that **does** hold up for local geometry |
|
| 378 |
+
| `matting` | alpha, the soft silhouette | crop-sized float16, 0 outside the mask | at very low alpha the **edge of the blade** appears, which is what the foil detector uses. A near-zero mean alpha means matting returned background for that crop |
|
| 379 |
+
| `pointmap` | the model's 3D point map, z channel | crop-sized float16, NaN outside | **kept but not used**: it orders the two fencers correctly in **76 %** of frames, while the y coordinate of the free foot, which costs nothing, gets it right **28 of 28**. It ships because it was asked for, not because it works |
|
| 380 |
+
| `foil / campo` | the threat cone as a graded probability field | see `campo_*` above | the fill **is** the weight function the measurement itself uses, so the figure and the number cannot disagree. The grey dashes are the forearm and are drawn even when the field is centred on the blade, so you can see how far the field moved off the axis. A dashed core arc plus `UNCERTAIN` means the weapon-hand cascade could not settle that fencer, and the front-foot vote is printed next to it |
|
| 381 |
|
| 382 |
### 3 - The movement events: one row per event, not per frame
|
| 383 |
|
|
|
|
| 397 |
| `worst_pose_provenance` | the worst pose provenance any frame of the event rests on | code, int | an event resting on an interpolated pose is weaker, and here you can see it |
|
| 398 |
| `gap_beyond_measured` | whether the event uses a pose from a gap longer than 24 frames | bool | flags data whose error is **not** verifiable |
|
| 399 |
| `side_valid` | whether that side really is a fencer | bool | in **6.1 %** of clips the delivered side is the **REFEREE**, and **all four** detection routes tried have failed. This flag is a **lower bound** |
|
| 400 |
+
| `weapon_hand` | which hand holds the weapon: A (root 21) or B (root 42) | category | the **cascade** scores **1.000 over 19 of 19 fencers** of a held-out set that played no part in choosing the rule, and 0.949 over all 59 with ground truth. Its front-foot rung alone covers **12 of those 19, with no failure**. What cannot be decided is flagged `uncertain` instead of guessed |
|
| 401 |
+
| `campo_mu_gr / campo_sigma_gr / campo_origen` | the blade probability field: centre direction, width, and where it came from | deg, deg, code | `origen 0` = the blade was **SEEN**, so the centre is the **measured** blade angle and the half-angle is **8.6 deg**; `origen 1` = the centre is the **assumed** forearm direction and the half-angle is **33.0 deg**. Both are the p90 of a measured residual, so seeing the blade narrows the field **3.8x** and centre and width are one decision. `temblor_hoja_gr` (4.2 deg) is a different quantity and ships separately: the frame-to-frame uncertainty of the blade's **direction**, not the width of the field. They never merge: **a measured 1 and an assumed 1 are not worth the same** |
|
| 402 |
|
| 403 |
|
| 404 |
---
|
|
|
|
| 436 |
controls count **clips**, some count **fencers** and some count **frames**, and pooling them would
|
| 437 |
invent a figure that was never measured.
|
| 438 |
|
| 439 |
+
Four results in this project fell apart when the sample grew or when a control was finally written -
|
| 440 |
+
15/15 became 22/23, a feature changed sign on a bigger set, a blade dispersion went from 14.0 deg to
|
| 441 |
+
4.2 deg, and a **published accuracy turned out to belong to the whole cascade and not to the rule it
|
| 442 |
+
was credited to**. That is why criteria here are declared before measuring, samples are broken down
|
| 443 |
+
by bout and by venue before being aggregated, and **accuracy and coverage are never merged into one
|
| 444 |
+
sentence**.
|
| 445 |
|
| 446 |
---
|
| 447 |
|