Swim Timing

Why this needed a trained model. A standard object detector run on race footage finds eleven people and every one is a spectator on the deck. It detects zero swimmers, because it has never seen a human who is mostly underwater in churning white water. EdgeSwim-3 was trained specifically for this.

What to expect. Trained on competition footage shot from the stands: wide, above water, side-on. It holds up on pools it has never seen, but it has two real blind spots. It finds nothing during a broadcast close-up, because a swimmer filling the frame looks nothing like its training data. And a swimmer is invisible while submerged off each wall, which no detector can fix. On your own pool with a fixed camera it does better than any of this — that footage is what a venue engagement produces.

See it work

A race the model has never seen, scored against the official annotation frame by frame. The rings are what EdgeSwim-3 tracked; the thin boxes are where the human annotator says each swimmer actually was.

Across the clip it tracks 70% of annotated swimmer-frames, and when it has a swimmer it sits a median of 14 pixels from the annotated centre on a 1920-wide frame. Lane 8 comes in at 100%, lanes 6 and 7 at 93%; lane 1 sits at the far edge and is missed entirely, which the panel shows rather than hides. Finish order and gaps are read from the official annotation, not from our own output.