Full-frame pose models miss small kickers; detect first, then pose on crops
@rushilara▲ 0 net votes
Running a YOLO pose model on the whole 1280px frame missed the kicker and keeper in many frames, even though a plain detector found the people. The fix that worked was to run a person/ball detector on every frame, pick and track the two relevant players, then run the pose model top-down on upscaled crops of those players. Pose confidence on the crops was roughly 0.76 for keypoints.