Research map / Video / Video segmentation and tracking
Precise Event Spotting in Sports Videos: Solving Long-Range Dependency and Class Imbalance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an end-to-end RegNetY spatio-temporal extractor with a novel ASTRM module, Bi-GRU temporal block, Soft-IC loss for class imbalance, and ASAM optimization for precise event spotting in sports videos.
Problem
Precise event spotting in long untrimmed sports videos requires capturing long-range temporal dependencies and handling imbalanced event class distributions, which existing frozen-feature methods handle poorly for rare classes.
Approach
Train a CNN spatio-temporal feature extractor end-to-end, enhanced by an Adaptive Spatio-Temporal Refinement Module (local spatial, local temporal, global dynamic-kernel temporal) and a Bi-GRU for long-range context, plus a Soft-IC loss that compacts class features under mixup labels, optimized with ASAM.
Key results
- On SoccerNet V2 tight setting, the proposed method reaches 73.74 mAP versus 71.33 for COMEDIAN and 61.19 for the E2E-Spot baseline, with 6.46M params.
- On Tennis, the method improves delta=0 mAP by 15.67% over the E2E-Spot baseline (61.01 vs 45.34).
- Ablation shows Soft-IC loss raises tight mAP from 72.65 (ASAM only) to 73.74 and helps minority classes such as Red card.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ASTRM to the E2E-Spot RegNet-Y 200MF baseline (ablation) | E2E-Spot on SoccerNet V2 test · SoccerNet V2 test, tight | mAP: 61.19 -> 68.63 | Helped |
| Add ASAM with ASTRM (ablation) | RegNet-Y 200MF on SoccerNet V2 test · SoccerNet V2 test, tight | mAP: 68.63 -> 72.65 | Helped |
| Replace ASAM with SAM (ablation) | RegNet-Y 200MF + ASTRM on SoccerNet V2 test · SoccerNet V2 test, tight | mAP: 72.65 -> 71.89 | Hurt |
| Add Soft-IC loss instead of Focal Loss (ablation) | RegNet-Y 200MF + ASTRM + ASAM on SoccerNet V2 test · SoccerNet V2 test, tight | mAP: 70.36 -> 73.74 | Helped |
| Replace Soft-IC loss with IC Loss (ablation) | RegNet-Y 200MF + ASTRM + ASAM on SoccerNet V2 test · SoccerNet V2 test, tight | mAP: 73.74 -> 72.05 | Hurt |
| Use all three ASTRM blocks instead of only local spatial (ablation) | Proposed method on SoccerNet V2 test · SoccerNet V2 test, tight | mAP: 69.91 -> 73.74 | Helped |
| Use all three ASTRM blocks instead of only local temporal (ablation) | Proposed method on SoccerNet V2 test · SoccerNet V2 test, tight | mAP: 70.41 -> 73.74 | Helped |
| Use all three ASTRM blocks instead of only global temporal (ablation) | Proposed method on SoccerNet V2 test · SoccerNet V2 test, tight | mAP: 71.65 -> 73.74 | Helped |
| Add the local temporal block to local spatial and global temporal blocks (ablation) | Proposed ASTRM on SoccerNet V2 test · SoccerNet V2 test, tight | mAP: 71.82 -> 73.74 | Helped |
| Add the global temporal block to local spatial and local temporal blocks (ablation) | Proposed ASTRM on SoccerNet V2 test · SoccerNet V2 test, tight | mAP: 71.06 -> 73.74 | Helped |
| Replace Bi-GRU with Bi-LSTM (ablation) | Proposed method on SoccerNet V2 test · SoccerNet V2 test, tight | mAP: 73.74 -> 68.15 | Hurt |
| Increase clip length from 100 to 128 frames (ablation) | Proposed method on SoccerNet V2 test · SoccerNet V2 test, tight | mAP: 71.20 -> 73.74 | Helped |
Surprising findings
- Focal loss reduced SoccerNet tight mAP (70.36) relative to the model without it, versus Soft-IC gains.
- Larger/more complex temporal modules (Transformer, Bi-LSTM, MS-TCN, deformable attention) underperformed the simple Bi-GRU.
Limitations
- Performance gains depend on input resolution and larger networks are costlier; authors note resolution trade-off.
- Soft-IC loss gives only a small overall improvement over ASAM alone on SoccerNet; main gains concentrate on minority classes.
Experimental setup
Datasets: SoccerNet V2 Tennis Figure Skating (FS-Comp, FS-Perf) FineGym
Benchmarks: SoccerNet V2 Tennis FS-Comp FS-Perf FineGym
Metrics: mAP tight mAP loose mAP at delta 0/1/2
Baselines: E2E-Spot T-DEED UGLF COMEDIAN Spivak ASTRA SpotFormer Baidu (TF) TSP + MS-TCN TSP + ASFormer 2D-VDP + MS-TCN VC-Spot
Builds on: E2E-Spot T-DEED Instance Contrastive loss ASAM mixup
Closest papers by meaning
- ViSTec: Video Modeling for Sports Technique Recognition and Tactical Analysis (AAAI 2024 · similarity 0.84)
- $F^3Set$: Towards Analyzing Fast, Frequent, and Fine-grained Events from Videos (ICLR 2025 · similarity 0.83)
- Towards Universal Soccer Video Understanding (CVPR 2025 · similarity 0.83)
- MomentSeeker: A Task-Oriented Benchmark For Long-Video Moment Retrieval (NeurIPS 2025 · similarity 0.83)
- PASS: Path-selective State Space Model for Event-based Recognition (NeurIPS 2025 · similarity 0.83)
- SportsHHI: A Dataset for Human-Human Interaction Detection in Sports Videos (CVPR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.