Introduction
Precise Event Spotting (PES) in sports analytics involves identifying the exact timestamp of specific game events, such as the moment a ball is struck or a foul occurs. Unlike general action recognition, PES requires high temporal accuracy. However, two significant challenges hinder current models: long-range temporal dependencies and extreme class imbalance. For instance, a "Shot on target" is often only verifiable several frames after the initial strike, requiring the model to incorporate future context. Furthermore, rare events like "Red Cards" are significantly underrepresented compared to common occurrences like "Ball out of play," leading to biased models.
Proposed Methodology
The researchers introduce an end-to-end framework that avoids the limitations of frozen, pre-trained features by training the entire pipeline on the target task. The architecture is built upon three primary innovations:
- Adaptive Spatio-Temporal Refinement Module (ASTRM): Integrated into a RegNetY backbone, this module employs parallel blocks for local spatial, local temporal, and global temporal refinement. This multi-faceted approach allows the network to capture fine-grained visual details while maintaining a broad temporal perspective.
- Bidirectional GRU (Bi-GRU): To address long-range dependencies, a Bi-GRU temporal block processes video chunks, allowing the model to leverage both preceding and succeeding frames to classify the current moment.
- Soft Instance Contrastive (Soft-IC) Loss: Designed to mitigate class imbalance, Soft-IC loss encourages the model to learn compact, well-separated feature representations for different event classes. It is specifically optimized to work with mixup augmentation, enhancing the model's robustness to rare samples.
The training process also utilizes Adaptive Sharpness-Aware Minimization (ASAM) to improve generalization by seeking flatter minima in the loss landscape.
Key Results and Discussion
The proposed method established new state-of-the-art (SOTA) benchmarks across several datasets, including SoccerNet V2, Tennis, Figure Skating, and FineGym. On SoccerNet V2, it achieved a 73.74 mAP in the strict "tight" setting, a 2.41% improvement over the previous best, COMEDIAN, despite having a more efficient parameter footprint.
A critical success of the model is its performance on minority classes. By utilizing Soft-IC loss, the model significantly improved detection for rare events. For example, performance on "Yellow-to-Red" card transitions jumped from approximately 10% mAP in baseline models to nearly 15% mAP.
Conclusion
This research demonstrates that an end-to-end trainable network, when equipped with specialized spatio-temporal refinement and contrastive loss, can effectively solve the dual challenges of temporal dependency and class imbalance in sports video analysis. Future work will focus on scaling the model to longer video contexts and higher spatial resolutions to further push the boundaries of temporal precision.
Summary updated