SoccerNet 2026: New AI spots ball actions per player with 58.94 F1
Two-stage pipeline with per-player attention boosts Macro-F1 by over 10 points.
Researchers Faisal Altawijri and Ismail Mathkour submitted a system to the SoccerNet 2026 Player-Centric Ball Action Spotting challenge, achieving a significant performance boost. Their approach is a two-stage pipeline: first, a Track-Aware Action Detector (TAAD) produces per-player action logits from broadcast video, enhanced with a temporal transformer for cross-frame context and several training improvements. Second, a Denoising Sequence Transduction (DST) transformer converts game-state features and TAAD logits into structured event sequences.
The key innovation is a two-stage per-player attention mechanism operating on game-state features. The researchers tested different attention orderings and found that spatial-first attention (cross-player attention before temporal attention) improves validation Macro-F1 by 1.87%. To maximize performance, they trained four diverse model variants and combined them using a Weighted Event Fusion ensemble with agreement filtering to suppress single-model false positives while preserving recall, including a special exception for the rare tackle class. The final system improved the challenge Macro-F1 from a baseline of 48.6 to 58.94, demonstrating the power of per-player modeling and ensemble strategies in sports video understanding.
- Two-stage pipeline: TAAD (Track-Aware Action Detector) with temporal transformer + DST (Denoising Sequence Transduction) transformer.
- Novel spatial-first per-player attention ordering yields 1.87% Macro-F1 improvement over alternative orderings.
- Ensemble of 4 model variants with Weighted Event Fusion and agreement filtering raises final Macro-F1 from 48.6 to 58.94.
Why It Matters
This system brings automated soccer analytics closer to human-level accuracy, enabling real-time per-player action insights for coaches and broadcasters.