DIME
Get the picksJoin

PROJECT · findings note

project_cv_rebound_tracking_spec

8-agent research+design workflow (2026-07-23) on a proposed broadcast-video CV player-tracking system — biggest finding is that NBA already has real public rebound-tracking data, no CV needed there.

← All findingssource · memory/project_cv_rebound_tracking_spec.md

User wants a system to catch "stretch player never crashes the boards but I trusted the rebound prop anyway" mistakes (Jeremy Grant example) — asked whether to build computer vision on recorded League Pass/YouTube broadcast footage to track player positioning. Ran an 8-agent research+design workflow (4 research agents + 3 candidate architectures + synthesis, ~650K tokens, ~16.5 min) before any code.

Headline finding — check this before writing any CV code: stats.nba.com/stats/leaguedashptstats?PtMeasureType=Rebounding (+ PaintTouch/ElbowTouch) is REAL Second-Spectrum optical-tracking data, publicly exposed today, with fields (OREB_CHANCES, OREB_CHANCE_PCT, contested/uncontested splits, PAINT_TOUCHES) that answer almost exactly "was this player ever in rebounding position" — for NBA. This repo already has a working stats.nba.com proxy pattern to clone (see reference_nba_summer_league_stats — the same x-nba-stats-origin/token header gotcha applies). This should ship as a proxy + methodology flag (days of work) before any CV investment, and may resolve most of the pain if it's mostly an NBA problem.

WNBA has no equivalent — verified against stats.wnba.com's own nav (box-score categories only) and SkillCorner (lists NBA/G-League/NCAA D1, not WNBA). Real Second Spectrum tracking exists in WNBA arenas since 2024 but is team/front-office-only, no public API. WNBA is the only league where a from-scratch CV build is actually justified.

Recommended phased path (not yet started, no code written):

  • Phase 0 (days): ship the NBA stats.nba.com rebound-proxy → SDS route → data-hydrator PlayerBundle.RimEngagement field → new stat == "rebounds" branch in diagnostic.go:356-412's computeMethodologyFlags (mirrors the existing stat == "points" Usage branch at diagnostic.go:283-285).
  • Phase 0.5 (<1 day): verify WNBA PBP is actually live — a WNBA PBP wiring commit landed on pbp-service's main only ~2 days before this research and was NEVER confirmed with a real ingested game (live-fetch verification was blocked from dev-machine IPs per the commit's own note). Hard prerequisite for Tier B, must be confirmed not assumed.
  • Phase 1 (1-2 weeks, cheap, kill-gated): build the event-anchored CV pipeline ONCE on ~10-15 NBA games, correlate its rim-proximity output against Tier A's real ground truth. Pure architecture validation — if it doesn't track real optical-tracking data where ground truth exists to check it, stop; it won't work blind on WNBA. This POC is throwaway, never wired into production for NBA (Tier A already covers NBA).
  • Phase 2 (conditional, realistically 6-10 weeks not the 4-8 proposals guessed — first-ever CV/ML-vision project for this team): port to WNBA only if Phase 1 clears + Phase 0.5 confirms PBP.
  • Phase 3 (months): promote from informational flag to hard LineEdges.Quality gate only after graded-pick validation, same governance already used for ActiveStatus/grade_v2/EB-shrinkage.

Tech choices for the CV tier (if built): YOLOv8/v11 + ByteTrack, event-anchored 3-10s windows only (not continuous tracking), per-window homography (not per-game), discard-on-camera-cut (no validated solution exists anywhere in the literature for bridging identity across broadcast cuts — even soccer's best-funded benchmark tops out at 63.81/100 GS-HOTA), jersey OCR with multi-frame majority vote constrained by roster-elimination from lineup+PBP data first. Explicitly rejected: pose/box-out detection (zero CV precedent found for this specific action), SAM2/segmentation (1-2 FPS on a T4, too slow), full-game ReID.

Real risks flagged, not hand-waved:

  1. Homography precision may not clear the bar needed — best published basketball-specific result is 126cm mean error, only 76.5% of frames under 1m — may not reliably distinguish "at the rim" from "just outside the paint," which is the exact distinction the feature needs.
  2. WNBA has no ground truth to validate against (unlike NBA) — a bad WNBA CV signal would only get caught the slow way, via graded-pick performance over a season.
  3. Zero CV/ML-vision expertise on this team (sports-scout-service is classical tabular ML) — every effort estimate should be read as optimistic.
  4. Legal/ToS: real but manageable gray zone (Betamax + NBA v. Motorola support the copyright side; ToS risk tops out at account termination, not litigation, no enforcement precedent found against private non-redistributed use) — mitigate by discarding raw footage post-processing, sourcing from paid League Pass over YouTube, staying internal-only. Operator should sign off knowingly.
  5. Open question for the operator, unresolved: is the actual pain point mostly NBA or WNBA? If mostly NBA, Phase 0 alone may fully resolve it. Also: WNBA already has a substantial higher-certainty backlog (see project_wnba_backlog) — worth an explicit call on whether CV beats clearing that queue for the next 2-3 months.

Why: Team has a track record of shipping sophisticated features that measured no lift and got reverted (see project_feature_experiments — USG%/position-aware DvP). Phase 1 is designed as a real kill-gate for exactly this reason, not a formality en route to a build already half-decided.

How to apply: Do not start CV implementation without re-confirming this plan with the user first — nothing has been built yet, this is a spec only. Full research detail (per-agent findings, all 3 candidate proposals) is in the workflow transcript, not duplicated here.