equine biomechanics

Equestrian coaching runs on expert eyes. A trainer watching a horse and rider sees asymmetries, rhythm breaks, and posture faults that are genuinely quantitative phenomena, measurable from video. But the field's measurement tools have historically been marker-based lab systems that don't survive contact with a real arena. The project goal: a pipeline that takes ordinary riding video and produces biomechanical analysis a working trainer would sign off on, running entirely on owned hardware.

Markerless pose estimation runs first: dual-subject inference (horse and rider simultaneously) built on DeepLabCut's SuperAnimal models, zero-shot, with per-keypoint confidence tracking and systematic handling of the failure modes zero-shot pose brings. Monocular depth lifting then recovers the third dimension from single-camera video, separating genuine biomechanical asymmetry from perspective artifact. Gait segmentation follows: stride cycles, footfall timing, and gait identification from keypoint trajectories.

From there the pipeline extracts kinematic features (joint angles, posture metrics, left/right symmetry, frequency-domain stability, and the coupling between horse and rider) and fits time-series models per subject. Those models produce prediction-vs-observation residuals: moments where the motion departs from its own recent dynamics. The residuals are the exogenous conditioning signal consumed by the REVEAL framework.

The scoring layer was developed in collaboration with a professional trainer, translating an expert's working rubric (hip and shoulder posture ranges, symmetry tolerances, rhythm expectations) into measurable thresholds, then validating those thresholds empirically against the pipeline's own measurements. The trainer's heuristics and the biomechanics literature disagreed in places; several of those discrepancies resolved into measurement-modality differences between 2D video and 3D motion capture, and the rest are being cross-checked against the PFERD motion-capture dataset.

This is the mission thesis in miniature: a domain the frontier models never trained on, an expert whose knowledge is the product, footage that shouldn't leave the operator's hands, and a pipeline that runs on consumer GPUs end to end. The structured signals it produces feed the corpus generation work, and the corpus feeds the training of small specialized vision-language models. Every stage runs locally and can be inspected.