Reliable per-player identity attribution in amateur volleyball video

Establish a reliable automated method for identifying which individual player made each volleyball contact in amateur single-camera footage, where jersey numbers may be optically illegible and multiple teammates may be crowded or occluded in the frame.

Background

The paper evaluates prompting and agentic vision-LLMs, trained computer-vision specialists, frozen video-world-model probes, and human annotation across the stages needed to convert amateur volleyball footage into player statistics. Although several stages can be recovered from volleyball’s structural rules—for example, rally outcomes from the next serving team—individual identity has no comparable rule-based shortcut.

The authors report that every automated paradigm struggles with identity attribution. The principal obstacles are low-resolution jersey numbers, unfavorable camera angles, occlusion, crowded scenes, and the fact that temporal context helps recognize actions but cannot reconstruct a jersey number that is never legibly visible. The paper therefore treats identity attribution as its central unresolved technical challenge.