Research map / Video / Action recognition
ViSTec: Video Modeling for Sports Technique Recognition and Tactical Analysis
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 ViSTec, a two-stage video model that segments table tennis strokes and classifies techniques using a learned transition graph over stroke sequences, evaluated on a 4000-clip WTT broadcast dataset.
Problem
Recognizing fine-grained stroke techniques in low-quality racket sports broadcast videos (occlusion, motion blur, small movements) to enable automatic tactical analysis, where standard action segmentation models fail.
Approach
Combine a stroke-event segmentation stage (cosine-peak targets on VideoMAE frame features) with a classifier conditioned on the previous stroke label and a learnable technique-transition graph updated with an entropy-weighted adaptive stride, trained jointly.
Key results
- ViSTec achieves F1@{10,25,50} of 79.3, 79.2, 78.5, accuracy 83.5 and edit 76.3, exceeding the best baseline MS-TCN (76.8, 74.8, 71.1; 78.2; 73.9).
- Removing the graph module drops ViSTec to F1@50 75.3 and accuracy 82.0, showing the transition graph contributes to performance.
- ViSTec runs at 39.3 FPS on a single A100, above typical broadcast frame rates.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the graph module (grh) (ablation) | ViSTec · WTT table tennis dataset | Accuracy: 82.0 -> 83.5 | Helped |
| add the graph module (grh) (ablation) | ViSTec · WTT table tennis dataset | F1@50: 75.3 -> 78.5 | Helped |
| use uncertainty term U in the graph update stride (ablation) | ViSTec · WTT table tennis dataset | Accuracy: 82.2 -> 83.5 | Helped |
| use uncertainty term U in the graph update stride (ablation) | ViSTec · WTT table tennis dataset | F1@50: 77.0 -> 78.5 | Helped |
Limitations
- Evaluated only on table tennis broadcast data from one dataset of 18 games
- Technique labels require expert annotation; dataset not released
- Graph conditioning uses only the previous stroke label, limiting context
Experimental setup
Datasets: WTT broadcast table tennis dataset (4000 rally clips from 18 games)
Benchmarks: WTT table tennis stroke dataset
Metrics: F1@10 F1@25 F1@50 frame-wise accuracy edit score
Baselines: C2F-TCN ASFormer UVAST SSTDA MS-TCN
Builds on: VideoMAE MS-TCN ASFormer UVAST
Closest papers by meaning
- SportR: A Benchmark for Multimodal Large Language Model Reasoning in Sports (ICLR 2026 · similarity 0.87)
- Towards Universal Soccer Video Understanding (CVPR 2025 · similarity 0.87)
- Learning Skill-Attributes for Transferable Assessment in Video (NeurIPS 2025 · similarity 0.86)
- FineSports: A Multi-person Hierarchical Sports Video Dataset for Fine-grained Action Understanding (CVPR 2024 · similarity 0.86)
- SoccerMaster: A Vision Foundation Model for Soccer Understanding (CVPR 2026 · similarity 0.85)
- LucidAction: A Hierarchical and Multi-model Dataset for Comprehensive Action Quality Assessment (NeurIPS 2024 · similarity 0.85)
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.