GameLens

Basketball film review with computer vision

Highlight reel possession

Mykeil Tzul game film · IG REAL Tape · 1.2–6.5 s

Original film

YOLOX tracking overlay

End-to-end transition

Mykeil Tzul game film · Broadcast excerpt · 3–9 s

Original film

YOLOX tracking overlay

Fast break

Mykeil Tzul game film · Broadcast excerpt · 4–10 s

Original film

YOLOX tracking overlay

Half-court set

Mykeil Tzul game film · Full broadcast clip · through the layup

Original film

YOLOX tracking overlay

Baseline camera

Mykeil Tzul game film · Vertical highlight reel · 0.5–7 s

Original film

YOLOX tracking overlay

Outdoor pickup benchmark

Evaluation footage · U.S. Army National Guard

Original film

YOLOX tracking overlay

Track IDs are temporary labels for person detections. They can change when players overlap.

From my playing days

Mykeil Tzul · 2020 NEBL highlights from Belize. Reviewing film as a player is what inspired GameLens.

Watch on YouTube · Personal highlights shown as published.

Watch my Senior Mixtape on YouTube

What worked, what missed

On three manually labeled frames from the outdoor pickup benchmark, GameLens matched 15 of 16 on-court player boxes at 50% overlap. There were 17 extra detections. Precision was 47% and recall was 94% on this small sample. Spectators and other people can be picked up by the detector. That evaluation measured detection boxes.

Manual labels · Evaluation output · GitHub source

Follow a jersey through the featured possession

I marked two jerseys from the opening to the end of the 5.3-second highlight. The audit samples every six frames, the final frame, and extra frames around the screen. It shows where overlapping boxes make identity hard to score.

JerseyVisible labelsUnique matchesMissedAmbiguousClear switchesChanges across overlap

Download jersey labels · Open per-frame results

Pickup footage by Sgt. Grayson Cavaliere, U.S. Army National Guard. Original footage