V-JEPA 2 Meets AQA: Phase-Based Scoring and Reference-Based Feedback for Mountain Biking

Authors: Santhiya Rajan, Savio Cardoz

Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)

Submitted: 27 September 2026 | arXiv:2610.30160 [cs.CV] | 27 pages, 10 figures

Overview

This study explores mountain-bike action quality assessment (AQA) under weak supervision, using frozen V-JEPA 2 video features to score jump technique and to propose candidate corrective feedback. As sports-video AQA increasingly leverages large pretrained video encoders in 2026, this work deliberately examines the limits of such a pipeline rather than claiming a deployable system.

Problem Setting

The training corpus comprises 95 field-recorded jump clips annotated only with binary labels denoting good or poor technique. These are coarse categorical judgments, not continuous expert quality scores; phase-boundary and aspect-level annotations are unavailable. This weak-label regime is representative of real-world coaching data, where detailed frame-level or dimension-level ratings are costly to obtain.

Method

The pipeline summarizes four heuristic movement phases using frozen V-JEPA 2 video features together with two-dimensional pose and motion measurements. A small transformer then predicts a binary-label score from these phase summaries.

A separate reference layer compares phase-specific measurements against those of clips labelled as good technique, generating candidate cues along three axes: phase, body parameter, and direction. The paper documents the measurements, network architecture, and evaluation protocol in detail so that the limits of each output remain explicit.

Results

In five-fold clip-level cross-validation, the full scorer achieved a pooled out-of-fold AUROC of 0.768 and average precision of 0.812. The authors describe these as exploratory figures, since the reporting fold also selected the neural training epoch and riders were not separated between folds.

Notably, removing the biomechanical scoring inputs raised AUROC to 0.809. A separate inner-selection run obtained AUROC 0.678, but used approximately 57 rather than 76 training clips per outer fold, so its gap from the test-selected run does not quantify selection bias. The unsupervised aspect head failed to learn distinct factors.

On a 17-clip external probe, an earlier checkpoint achieved binary accuracy of 0.412 and a Spearman association of −0.135 with expert scores. Candidate cues reached precision of 0.118 and recall of 1.0 against each of two coaches, with substantial disagreement about direction between them.

Interpretation

These findings do not establish rider-independent scoring or reliable corrective feedback. Instead, they map out the empirical work—rider-disjoint evaluation, stronger phase annotation, and multi-coach ground truth—needed before either application can be validated. The study is best read as a candid baseline for weakly supervised, encoder-driven AQA in sport.

Cite as: arXiv:2610.30160 [cs.CV] · DOI: https://doi.org/10.48550/arXiv.2610.30160

via ArXiv CV