CVPR 2026 model predicts next 5 seconds of soccer ball actions

Parthsarthi Rawat presents a hierarchical model for ball action anticipation in football broadcast video. The system processes a 30-second observation window divided into 5-second sub-windows. A shared local Transformer encodes clip-level features within each sub-window, then a GRU aggregates temporal context across all windows. Finally, a Transformer decoder with K input-conditioned event slots decodes prediction targets via three decoupled heads: objectness, class, and temporal offset. This design allows the model to anticipate ball actions (e.g., pass, shot, dribble) occurring in the subsequent 5-second window.

Two key innovations boost performance on rare action classes. Frequency-reweighted Hungarian matching systematically favors actions that appear less often in training data, and Gaussian soft targets provide smoother supervision for temporal bin regression. On the SoccerNet Ball Action Anticipation benchmark, the method achieves 17.91% mean average precision (mAP) on the test server, earning Rank 4 at the CVPR 2026 challenge. This work demonstrates that hierarchical processing with specialized loss functions can effectively anticipate fine-grained sports actions from broadcast footage, enabling real-time analytics and automated highlight generation.