← All projects

Sports Analytics · Switzerland

CrackFinder — equestrian analysis

A modular ML pipeline for equestrian parcours analysis — it detects jumps, segments the horse with SAM3, and predicts the horse's level, end-to-end from a single raw competition video.

ROLEML Engineer CLIENTConfidential · Switzerland STATUSActive
Horse mid-jump with SAM3 segmentation mask outline

The challenge

From raw video to a graded level

A full parcours recording has to become one meaningful horse-level prediction.

Find the jumps first

The signal lives in the jump moments — they must be located and clipped out of long footage.

Isolate the horse

Level cues are in the horse's motion, so it must be segmented cleanly from a busy arena.

The approach

01

Jump detector

Detects jump events and extracts individual jump clips from full parcours videos.

02

Horse segmentor (SAM3)

Isolates the horse ROI in each jump clip with SAM3-based tracking and segmentation.

03

Level predictor

Predicts the horse's level from sampled segmented jump clips.

04

Modular pipeline & MLOps

A Fire CLI, DVC-tracked data and models, and checkpoint export/evaluation tie it together.

Outcomes

3 stages

Modular pipeline

Jump detection → segmentation → level prediction.

End-to-end

Video → level

Raw parcours clip to predicted horse level in one command.

MLOps-ready

CLI + DVC

Fire CLI, DVC versioning, and checkpoint export flows.

Stack PyTorch SAM3 Fire CLI DVC Segmentation Video pipelines Python 3.12