Datasets:

EXCALORIO
/

fencingly-datos

Languages:

Tags:

License:

Dataset card Files Files and versions

xet

Community

The Dataset Viewer has been disabled on this dataset.

FENCINGLY

What we extract from a fencing clip, what each piece is worth, and the videos that show it

We are trying to learn who has right of way in a foil action. Foil is a weapon of convention: the valid touch is not simply the first one to land, so the answer cannot be read off the scoring lights. That is the whole reason this project exists - and it is the one thing that does not work yet: our detector scores 69.7 % against 73.8 % for never answering at all, so it ships off. It is in the capability table below with that number next to it, and this page is about everything that is extracted along the way.

Everything below is generated by us from the clips: skeletons, masks, normals, alphas, threat cones, piste maps. Every video and every figure on this page comes out of the pipeline - 38 videos in all - and each one carries its measured error printed on the frame, not just its name.

Credit and thanks

The clips come from FencingDatabase - https://fencingdatabase.com/

Thank you to the FencingDatabase team for letting us use their clips. Without that corpus none of what follows exists: it is the only place where thousands of foil touches are collected, cut and searchable, and building it from scratch was never an option for us. If you work on fencing video, start there: fencingdatabase.com.

The whole thing in one frame

Top left, the two skeletons on the clip with each fencer's position in metres. Under them, the label with what each fencer is doing right now - march, retreat, lunge, halt. On the right, the trace: position along the piste against time. Along the bottom, the piste from above, to scale, with each fencer's trail behind them and the events appearing as they happen.

10 s, generated over the clip. Everything below is one layer at a time, with what it is worth next to it.

What is extracted, layer by layer

Every video below is the same clip, one layer at a time, at the clip's own frame rate. The label on each frame says what you are looking at and what it measured.

How a pose gets found: four rounds, and what each one recovers

A pose is not simply detected or not. There are four rounds, and each one is counted separately so you can see what would be missing without it.

The pose is Sapiens 2 in every round. What changes between rounds is where the box comes from: YOLO gives only the box, and the 308 points are always measured by the same pose head. Round 2 exists because that head is top-down - it needs no detector to confirm the person, only a box - so it can try exactly where the detector already failed twice. Round 3 is the only one with no measurement left in it.

round	where the box comes from (the pose is always Sapiens 2)	recovers	of 21,950 fencer-frames
0	YOLO on the whole frame	21,274	96.92 %
1	YOLO again, on a crop - the redetection	252	1.15 %
2	interpolated box, no detector at all	50	0.23 %
3	no box and no pose: the pose itself is interpolated	600	2.73 %

Read it as a cascade: without round 1 the pose would be missing in 3.08 % of fencer-frames. Round 1 - the redetection - recovers 252 of them, leaving 2.96 % with no measured pose. Round 2 recovers 50 more, and what is left, 2.73 %, is interpolated and carries a flag saying so.

Round 2 is only attempted where the gap is 24 frames or shorter (50 of 295 cases). Beyond that it is not a tuning problem: the quality gate passes 0.6 % of the time with a fencer in the box and 12.7 % on an empty patch of background, so it does not discriminate at all - the fencer is simply not in those frames.

A redetection you can look at. Left: the full frame, with the box where the detector found nobody. Right: the same box as a crop, where the pose came out.

1 - POSE: 308 keypoints per fencer

Two fencers, 308 points each, every frame. Green is side 0 (image left), orange is side 1. A red box is a person the detector found and we did not select. The insets at the bottom are the weapon hand at native resolution, because each hand is 21 of those 308 points and at full frame they are a blur.

8 s of the layer, 720 px. The MP4 below is the whole clip at full resolution - the GIF is here because GitHub does not play the video.

pose_BabyishIdioticCockerspaniel.mp4 - MP4, 3.1 MB

pose_proudDungBeetleBoglenTerrier.mp4 - MP4, 3.6 MB

More clips, same layer: uptightSlothEveningBat

Full-resolution stills:

2.26 px torso error from running the pose at 384x288 instead of 1024, which is 7.8x cheaper
96.0 % of delivered poses are measured; the rest are interpolated and say so
0 held poses: a held pose is the same pose repeated, and it would look like a fencer standing still
10,140 keypoints were landing outside their own person's box - on the other fencer - and are now zeroed
there is no wrist index in this map: the wrist is the hand root. The map was measured, not read off a datasheet
the invented poses can be partly replaced by measured ones, and this was just measured. Of the 650 invented poses in 38 clips (2.96 % of delivered fencer-frames, all of them interpolated - held is 0 everywhere), re-running the pose head alone on that exact frame recovers a valid measured pose in 50 of the 295 whose gap is <= 24 frames. In a control where fully measured frames were hidden and re-measured, the rescue costs 1.94 % of torso length against 21.02 % for interpolating, and it is closer to the truth in 98.2 % of 626 cases
and for gaps of 25 frames or more it does not work, which is the useful half of that result: only 2 of 355 pass the quality gate, while an empty patch of background in the same frame passes 45 of 355. The rescue finds a fencer in the estimated box less often than in a piece of floor, because the fencer is not in those frames at all. So beyond 24 frames there is no measurement is not a tuning problem, and those 355 - 54.6 % of everything invented - stay unverifiable

2 - SEG: body-part segmentation

The cleaned mask, blended only inside itself. Outside it the frame is just darkened, which is deliberate: that is how you can check the mask landed on the right person instead of on the referee walking past.

8 s of the layer, 720 px. The MP4 below is the whole clip at full resolution - the GIF is here because GitHub does not play the video.

seg_BabyishIdioticCockerspaniel.mp4 - MP4, 2.1 MB

seg_proudDungBeetleBoglenTerrier.mp4 - MP4, 2.4 MB

More clips, same layer: uptightSlothEveningBat

Full-resolution stills:

referee leakage into the mask measured 0.0000
the mask is (matting >= 0.5 OR seg != background) AND nearest-skeleton AND box guard, and the loss is attributed to the three of them separately - a low recall does not say what to fix
stored per crop, not per frame: the two masks cover 4.6-4.8 % of the frame, so full-frame storage would waste 21x

3 - NORMAL: surface normals

One axis per colour channel. This is the dense layer that does hold up for local geometry.

8 s of the layer, 720 px. The MP4 below is the whole clip at full resolution - the GIF is here because GitHub does not play the video.

normal_BabyishIdioticCockerspaniel.mp4 - MP4, 2.1 MB

normal_proudDungBeetleBoglenTerrier.mp4 - MP4, 2.4 MB

More clips, same layer: uptightSlothEveningBat

Full-resolution stills:

outside the mask the value is NaN and not 0, because the zero vector is not a unit normal but it would average silently inside a convolution

4 - MATTING: alpha, the soft silhouette

The soft silhouette. At very low alpha the edge of the blade appears, and that is what the foil detector works on.

8 s of the layer, 720 px. The MP4 below is the whole clip at full resolution - the GIF is here because GitHub does not play the video.

matting_BabyishIdioticCockerspaniel.mp4 - MP4, 1.9 MB

matting_proudDungBeetleBoglenTerrier.mp4 - MP4, 2.2 MB

More clips, same layer: uptightSlothEveningBat

Full-resolution stills:

the footer prints the mean alpha per side. A near-zero one is not a bug in the video: matting returns background when a second person is inside the crop, and then that fencer's silhouette is empty. A null has to say it is a null and why

5 - POINTMAP: the depth-like channel we do NOT use

Shown because it was asked for, not because it is used. It is kept as a channel and the pipeline ignores it.

8 s of the layer, 720 px. The MP4 below is the whole clip at full resolution - the GIF is here because GitHub does not play the video.

pointmap_BabyishIdioticCockerspaniel.mp4 - MP4, 2.2 MB

pointmap_proudDungBeetleBoglenTerrier.mp4 - MP4, 2.6 MB

More clips, same layer: uptightSlothEveningBat

Full-resolution stills:

it orders the two fencers in depth correctly in 76 % of frames - and the y coordinate of the free foot, which costs nothing, gets it right 100 % (28 of 28)
Sapiens 1 depth scored 73.2 %, below the free baseline, on a newer model's predecessor
the 3D cone is off for the same reason: negative Z in 11.4 % of one clip, implied field of view 21-144 deg where a fencing camera uses 8-30, and the 11 places where 2D and 3D disagreed were noise - permuting the depths at random reproduced them exactly

6 - FOIL: the threat cone, as a graded probability field

The cone is what turns where the arm points into who can be hit. It is not a flat wedge: the fill is the weight function the measurement itself uses, so the figure and the number cannot disagree. This project already lost a day to a visualisation that did not match its own data, so the drawing and the metric are now the same code.

The yellow cross is where the two cones cross each other. That is the test - cone against cone, not cone against torso.

8 s of the cone. The MP4 below is the whole clip at full resolution - the GIF is here because GitHub does not play the video.

foil_BabyishIdioticCockerspaniel.mp4 - MP4, 2.7 MB

foil_proudDungBeetleBoglenTerrier.mp4 - MP4, 3.7 MB

More clips, same layer: uptightSlothEveningBat

Full-resolution stills:

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
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 %
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
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
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
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
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

The piste: where each fencer is, in metres

This was rebuilt from scratch, because the first version had a defect worth stating plainly: the origin was a convention, not a measurement, so fencers appeared to start near the centre line when they do not. It now rests only on what the rulebook fixes, which is what makes it independent of the piste model.

Left: the 308-point skeleton with each fencer's position in metres. Right: the trace - position along the piste against time, which is the racing-line view. Bottom: the piste from above, to scale, with the regulation lines and the events appearing as they happen. The counter is there so the count can be judged frame by frame rather than taken on trust.

12 s of the whole four-zone frame: skeletons and metres top-left, the label with the event underneath, the trace on the right, the piste from above along the bottom.

The same thing zoomed into the two panels that carry the measurement, because at full frame their axis labels are too small to read:

The trace. Position along the piste on the horizontal axis, time downwards. Green is the left fencer, orange the right. Both coordinates in this panel are measured: x by homography, time exactly.

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.

trace_AbandonedBiodegradableCoqui.mp4 - MP4, 1.0 MB

trace_AdorableVeneratedDuiker.mp4 - MP4, 5.7 MB

trace_AliveIgnorantBarebirdbat.mp4 - MP4, 9.1 MB

trace_BountifulSilentEarwig.mp4 - MP4, 9.0 MB

trace_CautiousBeneficialLacewing.mp4 - MP4, 7.4 MB

trace_ForcefulAntiqueAlbacoretuna.mp4 - MP4, 8.9 MB

trace_GentleQuerulousElephantseal.mp4 - MP4, 5.6 MB

trace_HairyCourteousEarwig.mp4 - MP4, 6.2 MB

trace_IlliterateDisloyalFulmar.mp4 - MP4, 7.7 MB

trace_InsecurePettyCreature.mp4 - MP4, 8.4 MB

trace_SeparateHoarseFeline.mp4 - MP4, 5.9 MB

trace_TestyWatchfulBarebirdbat.mp4 - MP4, 5.4 MB

trace_UglyNiceAplomadofalcon.mp4 - MP4, 6.1 MB

trace_VibrantJoyousAsiaticgreaterfreshwaterclam.mp4 - MP4, 5.5 MB

trace_WastefulYawningAmericanquarterhorse.mp4 - MP4, 2.8 MB

trace_WholeTallAlpineroadguidetigerbeetle.mp4 - MP4, 7.9 MB

At quarter speed, for judging the detection lag frame by frame: AliveIgnorantBarebirdbat | ForcefulAntiqueAlbacoretuna | GentleQuerulousElephantseal | UglyNiceAplomadofalcon

What is measured now, each with its own denominator, never pooled:

	value	out of
rectified strip obtained (both edges)	20	22 clips
origin MEASURED from the t.19 colour frontiers	13	22 clips
of those, unambiguous	11	22 clips
of those, with redundancy, i.e. controlled	5	22 clips
both fencers land on the piste (\|x\| <= 7.5 m)	11	the 13 anchored clips
held-out landmark prediction error	p50 5 cm, max 8 cm	5 clips
lateral travel <= 0.60 m (the physical control)	11	40 fencer-clips
clips where both fencers pass that control	3	22 clips
lateral travel p50 / p90 / max	0.75 / 1.63 / 3.33 m	40 fencer-clips
foot contacts falling outside the piste	p50 2.7 %, max 23 %	20 clips
fencers crossing each other (physical control)	0	20 clips
extra lateral plank detected	0	22 clips

The 5 cm row is the one that matters: fit the anchor with two landmarks, predict a third one it never saw, and measure the miss. That is the independent control this part never had.

the origin is measured, the along-piste scale is not. Three independent routes were tried and none passes its own control: painted lines by separation ratios (fits 14 of 20 clips, ambiguous in 10 of those 14, off by 1.3-2.6x against the reference), square pixels from the two ground vanishing directions (an exact camera solution that re-projects fencer heights of 0.4 to 4.9 m), and colour zones (13 of 22, only 5 controlled). There is a geometric reason it is not a matter of tuning: two parallel lines with a known separation fix the scale across the piste everywhere and say nothing about the scale along it. So the map is labelled origin MEASURED, scale NOT verified
across the piste is now metres between the two detected edges, bounded by construction. The earlier version assigned each fencer a fixed lane, so the one on the right always appeared glued to the right edge; that is gone
the lateral coordinate is still shape without scale: travel of p50 0.75 m against a control of < 0.60 m, and in only 3 of 22 clips do both fencers pass it. There is a physical reason: across the piste is the camera's depth direction, and a foot in the air projects displaced towards the background
there is more than one piste model, and here is the number: of the 13 clips where the final zone is identified, 8 are darker than the piste and 5 are lighter. A detector that looked for grey would fail on more than a third of the corpus, so everything here detects a colour change and never a colour
where there is no scale, the metric fields come out None, not zero. Writing zero would claim the fencer was standing on the centre line

The movement events

March, retreat, lunge, halt, arm extension - one row per event, not per frame. This is the low-dimensional series the model is meant to learn from.

90 % recall, and the confusion that mattered, march mistaken for lunge, is 0
the discriminator is the rear foot: in a lunge the front foot flies and the rear foot stays planted. That binary property, not the distance, is what separates the two, and the ratio is measured in the extension phase and not over the whole window
overlaps between different event types went from 73 to 0, enforced by an assertion
detection lag +0.5 frames (+17 ms) over 62 pairings, with no bias
balestra and jump ship with 0 events. The balestra was killed by ballistics: the 8.5 cm of rise it needs would require 7.9 frames of flight, and the candidate events lasted one frame

Which hand holds the weapon

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.

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
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
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
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
what cannot be decided is flagged uncertain rather than guessed

The open defect: sometimes one of the two fencers is the referee

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:

route	why it failed
uniform whiteness	the gold lame fires it on correct pairs; paired separation came out at -0.002, i.e. backwards
mask held in hand	with a loose threshold it becomes a darkness detector: fires on 469 of 3,386 patches while only 1 of 39 fencers holds a mask
geometry	the referee passed all three geometric gates without touching them (separation 1.93 of 3.20, foot delta 0.46 of 0.55, feet-in-band 0.894 of 0.40) - and it was predictable, because a referee walks beside the piste
movement behaviour	the referee scored 0.524 events per second, above the p10 of real fencers (0.493) and four times its own opponent's rate: walking arm-swing fires the arm-extension event

Every row carries a side_valid flag and it is a lower bound. The reason this is hard is worth stating: the referee's pose is perfectly good. There is nothing wrong with the pose; what is wrong is whose it is. No pose-quality metric will ever catch it.

Where each capability stands

Several of these are published as not working. That is deliberate: a field without its error is a promise, and a capability without its failures is a brochure.

capability	state
both fencers tracked, no intruders	works - median 100 % of frames, worst clip 91 %
always two poses, invented ones marked	works - 0 delivered frames without both
provenance of every point	works - and it is the part that cannot be recovered later
the five model channels	works - referee leakage 0.0000
movement events	works - 90 % recall, `march -> lunge` confusion = 0
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
piste rectified from its two edges	works - 20 of 22 clips
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
scoring lights	works - 90.5 % with a per-venue region of interest
foil blade and threat cone	partial - the blade is found in 42.6 % of 1080p crops but 1.6 % at 720p
position across the piste	partial - metres between detected edges, but the physical control passes in only 3 of 22 clips
scale along the piste	does NOT work - three independent routes, none passes its own control, and there is a geometric reason
referee told apart from a fencer	does NOT work - ~6 % of clips, four routes failed
right of way (the goal)	does NOT work - 69.7 % against 73.8 % for never answering, and beaten on both halves of the job
parry	not implemented - 0 annotated, blade contact not verifiable by eye in this material
depth from `pointmap`	discarded with measurements - the free `foot y` orders depth better

Parry - why there is nothing here

0 parries annotated, because blade contact turned out not verifiable by eye in this material: of 16 blind plates, zero were decidable. Without ground truth there is nothing to measure against, so there is nothing to ship. halt is not a parry: it means the fencer stopped moving.

Scale of the data


clips in the corpus	8,559
clips with 308-point poses today	732
movement events extracted	1,717 over 65 clips
frames per clip, median	307
resolution	81 % at 1280x720, 18 % at 1920x1080, 0.8 % at 854x480
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
over the whole corpus, on one GPU	41 h and 67 h

The data table

1 - What comes out per clip, and in what shape

T is the number of frames in the clip; the 2 is the two fencers, side 0 on the image left and side 1 on the image right.

field	shape on disk	what it is	format	measured precision
`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
`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
`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
`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
`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
`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
`pres`	`(300, 2)`	whether that fencer's pose in that frame is MEASURED	bool	0 delivered frames without both fencers, across 28 clips
`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
`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
`filtro_fuera_de_caja`	`(), in 21 of 732 files`	keypoints zeroed for falling outside their own box	count, int32	10,140 across 28 clips
`recortados_por_pose`	`(), in 21 of 732 files`	frames trimmed off the ends for not having both poses	count, int32	80 across 22 clips, which is what leaves 0 held poses
`W`	`()`	video width	px, int	81 % of the corpus is 1280x720
`H`	`()`	video height	px, int	18 % is 1920x1080, 0.8 % is 854x480
`T`	`()`	frames in the clip	frames, int	median 307. The previous pipeline stored a 129-frame window, not the whole clip

2 - The five model channels kept per fencer

channel	what it is	format	why it is there, measured
`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
`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
`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
`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
`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
`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

3 - The movement events: one row per event, not per frame

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.

column	what it is	format	measured precision
`clip / side`	which clip and which fencer (0 = image left, 1 = image right)	str / int	side 0/1 is an image convention, not left/right-handedness
`type`	march, retreat, lunge, halt, arm extension, action, jump	category	confusion matrix against gestures annotated by eye: 5 of 6, with `march -> lunge` = 0. `jump` ships with 0 events: not detectable here, and a ballistic assertion enforces it
`index_of_type`	the event's number within its own type and side	int	this is what makes the third march of the right-hand fencer a one-line query, and it is the same number the video counter shows
`frame_start / frame_end`	when the event starts and ends	frame, int	detection lag measured at +0.5 frames (+17 ms) over 62 pairings, with no bias. The boundary error is still unmeasured: that is what is being annotated now
`magnitude_m`	how far the fencer moved during the event	m, float	march p50 0.346 m, lunge p50 0.565 m. Depends on the piste scale, which is not verified along the piste
`x_start_m / x_end_m`	position along the piste	m from the centre line, float	the origin is measured from the t.19 colour frontiers in 13 of 22 clips, and a held-out landmark is predicted to p50 5 cm. The scale along the piste is not verified. Per-frame tremor p95 0.025 m, against 0.13-0.20 m/frame for a real lunge. Convention: x grows towards image right
`y_start_m`	position across the width of the piste	m, float	metres between the two detected edges, bounded by construction, so no fencer is pinned to an edge. But the physical control (travel <= 0.60 m) passes for 11 of 40 fencer-clips, and in only 3 of 22 clips do both fencers pass it
`no_homography`	whether this clip has no metric scale at all	bool	6 in 10 clips have none. When true the metric fields come out None, not zero
`razon_pie_atras`	rear-foot displacement over front-foot displacement	ratio, float	the march/lunge discriminator: boundary 0.45, measured in the extension phase and not over the whole window
`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
`gap_beyond_measured`	whether the event uses a pose from a gap longer than 24 frames	bool	flags data whose error is not verifiable
`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
`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
`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

This table is the training input

Everything above is the model's input. Not the video: the series.

Per clip and per fencer the model gets the 308-point pose with its provenance, the five channels, the position along the piste where there is a scale, and the movement events as a row each - with their type, their index, their start and end frame, their magnitude in metres, and the flags that say how far each one can be trusted. The label to learn is right of way.

Which is why every row carries its precision and not just its name. Several are published as not usable yet, and one - the referee - as a known defect with a flag, because a model trained on data that hides its own defects learns the defects.

ev.query("side == 1 and type == 'march' and index_of_type == 3")

returns the third march of the right-hand fencer: start frame, end frame, second, metres covered, how long it took, and how much of it was measured rather than filled in.

Honest limits

Figures come from 19 to 60 clips depending on the measurement, not from the full corpus.

Ground truth for the weapon hand, for step counting and for blade contact was built by eye inside this project, not by an FIE referee, and is kept in files so it can be contradicted.

Where a number appears twice on this page with two different denominators, that is deliberate: some controls count clips, some count fencers and some count frames, and pooling them would invent a figure that was never measured.

Four results in this project fell apart when the sample grew or when a control was finally written - 15/15 became 22/23, a feature changed sign on a bigger set, a blade dispersion went from 14.0 deg to 4.2 deg, and a published accuracy turned out to belong to the whole cascade and not to the rule it was credited to. That is why criteria here are declared before measuring, samples are broken down by bout and by venue before being aggregated, and accuracy and coverage are never merged into one sentence.

Fencing terms

term	meaning
march / retreat	a step forward / a step backward
lunge	the explosive attack: the front foot flies and the rear foot stays planted. That binary property, not the distance, is what separates it from a march
halt	the fencer stops moving. This is not a parry, and the parry is not implemented
right of way	which fencer has priority for the touch
valid target	in foil, the torso. A white light means the tip landed off-target
on-guard	the basic stance; the front foot is on the same side as the weapon arm
lame	the conductive jacket that makes a valid touch register

Source clips: FencingDatabase - thanks to their team.

The extraction pipeline lives in a private repo (EXCALORIO/fencingly-extraccion, access on request).

Mirror of this page, same README and same videos: github.com/EXCALOFRIO/FENCINGLY.AI

Downloads last month: 73

Total file size:

291 MB