Research map / Video / Video segmentation and tracking
What Will Happen Next? Forecasting Player Moves in Sports Videos
Panna Felsen, Pulkit Agrawal, Jitendra Malik
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add nearest-defender position and distance features to the player features (F[3..9] vs F[8..9]) (ablation) | Random forest on overhead view · Water polo next-possessor (test, detected heads) | accuracy (%): 35.2 -> 44.1 | Improved |
| add ball possessor coordinates F[1,2] to the features (F[1..9] vs F[3..9]) (ablation) | Random forest on overhead view · Water polo next-possessor (test, ground truth heads) | accuracy (%): 48.8 -> 47.1 | No clear effect |
| use estimated detections instead of ground truth heads (ablation) | Random forest F[1..9] · Water polo next-possessor (test) | accuracy (%): 47.1 -> 45.5 | No clear effect |
| make the FCN also output player locations in addition to the next possessor (ablation) | FCN on raw images · Water polo next-possessor (test) | accuracy (%): 10% to 38.1% | Improved |
| use overhead-view random forest instead of FCN on raw images | Water polo next-possessor · Water polo next-possessor (test, detected heads) | accuracy (%): 35.2 -> 45.5 | Improved |
| use early fusion of 5 past frames instead of CNN + LSTM late fusion (ablation) | Basketball ball position CNN · Basketball ball position 1s ahead | mean distance error (% court length): 11.4 -> 10.8 | Improved |
| use early fusion of 5 past frames instead of CNN + LSTM late fusion (ablation) | Basketball ball position CNN · Basketball ball position 2s ahead | mean distance error (% court length): 17.1 -> 16.8 | Improved |
| use overhead-view CNN with hand features instead of raw-image CNN | Basketball event forecasting · Basketball event forecasting 1s ahead (full test split A) | mAP (%): 11.9 -> 22.6 | Improved |
| use random forest on 92 hand-designed features instead of Overhead CNN | Basketball event forecasting · Basketball event forecasting 1s ahead (full test split A) | mAP (%): 22.6 -> 26.4 | Improved |
| use random forest instead of Overhead CNN at 2s horizon | Basketball event forecasting · Basketball event forecasting 2s ahead (full test split A) | mAP (%): 13.7 -> 18.7 | Improved |
| train random forest on basketball and test on water polo (transfer) | Random forest F[1..9] · Water polo next-possessor (test) | accuracy (%): 45.5 -> 36.8 | Hurt |
Closest papers by meaning
- Where Will Players Move Next? Dynamic Graphs and Hierarchical Fusion for Movement Forecasting in Badminton (AAAI 2023 · similarity 0.86)
- LATTE-MV: Learning to Anticipate Table Tennis Hits from Monocular Videos (CVPR 2025 · similarity 0.85)
- Towards Universal Soccer Video Understanding (CVPR 2025 · similarity 0.84)
- Event2Tracking: Reconstructing Multi-Agent Soccer Trajectories Using Long-Term Multimodal Context (AAAI 2025 · similarity 0.84)
- SportR: A Benchmark for Multimodal Large Language Model Reasoning in Sports (ICLR 2026 · similarity 0.83)
- ViSTec: Video Modeling for Sports Technique Recognition and Tactical Analysis (AAAI 2024 · similarity 0.83)
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.
