Event-Driven Highlight Generation in Football Broadcasts: A Context-Aware Temporal Framework for Six-Class Action Spotting

This paper presents an end-to-end system for the automatic summarization of football matches through temporally accurate event spotting across six action categories: goal, card, substitution, kick-off, direct free-kick, and corner. These six classes were selected from the SoccerNet-v2 annotation taxonomy as a non-standard evaluation subset: the three standard SoccerNet-v2 classes (goal, card, substitution) are retained and extended with three additional context-dependent set-piece classes (kick-off, direct free-kick, and corner) that represent the primary unresolved challenge in temporal action spotting. We developed and empirically evaluated a framework based on a Temporal Convolutional Network (TCN) with capsule-based relational encoding applied to class-specific Time-Shift Encoded (TSE) labels using a composite segmentation-detection loss. Evaluated on the SoccerNet-v2 benchmark (300/100/100 train/validation/test splits) using the official PCA-compressed ResNet-152 features sampled at 2 FPS, the implemented TCN-Capsule baseline achieved a test-set mean Average Precision (mAP) of 58.16% at a 10 s tolerance threshold (δ = 20 frames), with a validation mAP of 58.55%, indicating strong generalization with only a 0.39 pp performance gap. The six per-class AP values underlying this result are Goal: 75.6%, Corner: 78.0%, Substitution: 56.9%, Card: 53.9%, Direct Free-kick: 52.8%, and Kick-off: 34.3%; the headline mAP is computed by the official SoccerNet area-under-precision-recall-curve script and is not a simple arithmetic average of these six values. In addition, five complementary evaluation metrics are introduced: Average Temporal Error (ATE), Temporal Recall@K (TR@K), Latency-to-Event (L2E), Temporal Causality Consistency (TCC), and Annotation Noise Sensitivity. The analysis reveals that visually salient events such as goals and corners are detected with high accuracy (AP > 75%, ATE < 5 s, AUC > 0.95 at frame-level), whereas context-dependent events, including kick-offs and direct free-kicks, remain challenging (AP < 55%, ATE > 8 s, AUC = 0.68–0.72). As future work, two architectural extensions are formally specified but not empirically evaluated: a Temporal Segment Network with Non-Local Blocks and a hybrid Transformer-CNN framework with optical flow fusion. Furthermore, an OpenCV-based highlight generation prototype is implemented; formal quality evaluation is planned as future work.