Performance · Published 26 August 2026
Benchmarking tennis video analysis without a vanity accuracy number
A reproducible protocol for calibration eligibility, event precision and recall, tracking quality, confidence, latency, and failed sessions.
Publish the corpus boundary
Stratify court surface, camera position, resolution, frame rate, lighting, player level, occlusion, rally length, and device. Separate the held-out set by session and venue so near-duplicate clips cannot leak across splits.
Record dataset fingerprint, model version, threshold, and excluded cases.
Measure events and relationships
Shot and bounce precision/recall are necessary but insufficient. Evaluate player association, event ordering, rally boundaries, source-frame offset, and court-reprojection error where calibration qualifies.
Report unavailable rate alongside accuracy; a model that abstains honestly behaves differently from one that guesses.
Measure the product path
Time upload authorization, transfer, queue wait, worker claim, normalization, each inference stage, artifact upload, canonical publication, notification, and report render. Publish video duration and GPU type.
Include failures, retries, duplicate claims, and deletion completion.
Compare equivalent modes
Live officiating, post-session coaching, manual tagging, and automated indexing are different jobs. Compare supported devices, input constraints, evidence linkage, methodology, privacy boundary, and price separately.
Do not infer a winner from a feature that one workflow does not attempt.
Production checklist
- Verify Owner isolation against the deployed environment, not a screenshot.
- Verify Atomic publication against the deployed environment, not a screenshot.
- Verify Measurement honesty against the deployed environment, not a screenshot.
- Verify Bounded video access against the deployed environment, not a screenshot.