Henry Equestrian Centers · a measuring project
We are building an instrument that measures a western dressage test, and right now that is the whole of it. Send a ride and it gets pulled apart frame by frame; you get the whole file back, including what the machine was unsure about and what it got wrong. Three rides traced frame by frame so far. It is free, it stays free, and nobody is being scored.
Hickory jogging the long side, traced off the video Tracee shot on a phone in a covered arena with cones for markers. Seventeen points on the horse and seventeen on the rider, five times a second, and nothing in the numbers underneath comes from anywhere except those points. The head, the hooves and the whole rider hold together. The barrel does not, and that is drawn honestly rather than smoothed over.
Every point above was found by a model that has never been shown a horse in tack, on footage shot on a phone. The head, the four hooves and the whole rider hold up; the trunk does not, which is why the numbers below use only the parts that do.
Every line above is computed rather than drawn, but only from the points themselves. There is no camera solve behind it: this was filmed hand-held, so the wireframe knows where the horse is in the FRAME and not where it is in the ARENA. That is the difference between the numbers under the video and the two rows the readout at the top marks not measurable, and it is a difference a tripod closes.
The bars are the share of 1,176 traced frames in which the model was sure enough about a joint to let it into a measurement — 0.50 confidence, which is a bar we set and can move. Nothing here is an average of averages or a score out of ten. It is the count, per joint, off one ride.
Read the left column against the right one. The right hip is good in 94 frames out of a hundred and the left hip in 48. The two hooves on the same leg are both above 88. That is not the model being unreliable — it is a camera standing on one side of a horse. The far shoulder, elbow, hip and stifle spend the test behind the barrel; the feet stay in the open underneath it. So the four hooves and the head are what the measurements are built on, and the barrel joints are drawn on the wireframe and used for nothing.
155.2 beats per minute across the jog, and a steadiness of 78.7 per cent. The bpm figure carries about ±15 per cent, because the onset envelope autocorrelates at a second lag as well; the steadiness figure is the sounder of the two. Both were checked against a control window over a standing halt, which returned 5 onsets against 58.
Rhythm is a different thing and we do not claim it here: it is the beat pattern itself — four beats to a walk, two to a jog, three to a lope — and the count, not the rate, is what a judge is reading.
Two, at 1:28 and 2:55, found by differencing frames at 4 fps with the camera shake subtracted first. Held six and seven seconds. Whether they were square is a different question and this footage cannot answer it — the horse is tail-on to the lens for one of them.
Centerline deviation, circle diameter, where a transition happened relative to the letter it was called at. All of it needs to know where the ground is, and the ground is found by fixing a camera and reading markers at known spots. Not one of the rides here carries a calibration, so not one of these numbers exists yet.
Topline, hind engagement, head carriage above the withers. The generic quadruped model put the tail head 35 pixels from the neck on an 845-pixel horse and was 80 per cent sure about it. Confidence and correctness came apart there, so nothing is built on those points and nothing is coloured to suggest otherwise.
Any western dressage test, any level, schooling or shown. It does not have to be a good ride — the ones where something went wrong are more use to us than the ones where nothing did. Six things make the difference between a ride we can describe and a ride we can measure, and five of them are free.
On a tripod, on a fence post, on a mounting block. Do not follow the horse, do not zoom. A camera that moves has to be tracked before anything else can be, and the two rides here that carry a note about the camera both say it panned and zoomed the whole way through.
Where the judge sits, and the higher the better — going from two metres to six does more for accuracy down the long side than doubling the resolution would.
Cones, buckets, plates, tape on the rail. Four visible markers at spots you can tell us is the whole difference between centimetres and adjectives. Twelve is better.
Hoofbeats are the tempo measurement. One ride so far came back with the hoofbeats under the noise floor, and nothing could be taken off it at all.
Head and tail of the file are fine — slates, tack checks, walking out. We trim it. Cuts inside the test are what we cannot work around.
That is the only way to find out where the machine disagrees with a judge, which is the thing we are actually trying to learn. It is not used to score you.