Hierarchical GRU with Input-Conditioned Slot Queries for Ball Action Anticipation
Researchers have developed a novel hierarchical model for anticipating ball actions in football broadcasts. The system utilizes a Transformer to encode clip-level features and a GRU to aggregate temporal context, predicting actions within a 5-second window based on a 30-second observation. This approach incorporates frequency-reweighted Hungarian matching to favor rare action classes and Gaussian soft targets for temporal supervision, achieving 17.91% mAP on the SoccerNet Ball Action Anticipation benchmark. AI
IMPACT This model advances AI's ability to predict events in real-time video, potentially impacting sports analytics and automated broadcasting.