Adaptive tiling prevents athletes from being cut off at tile edges in images.

Pinpointing an athlete's real-world spot from one broadcast image is tough because people appear at wildly different sizes. To fix this, researchers built a top-down system that uses a single calibrated frame. High-resolution photos are usually sliced into smaller tiles, which often cuts players in half. The team solved this with Boundary-Aware Adaptive Tiling. This tool uses rough box predictions to stretch tile borders until each athlete fits completely inside. Next, an adapted pose network finds two key points: the player's pelvis and where they touch the ground. The system then projects that ground spot into real-world coordinates. This approach cut down missing detections and showed that perspective distortion causes most remaining errors. It beat the baseline system by over 21 %.

Why it matters

Sports broadcasters and analysts need to track athletes accurately for game analysis and viewer overlays. Extreme size differences in high-resolution footage cause errors, but this method adapts how the image is split to improve reliability.

Questions answered