Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Event2Tracking: Reconstructing Multi-Agent Soccer Trajectories Using Long-Term Multimodal Context
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 Event2Tracking, a Transformer that fuses long-term (up to 60 s) noisy broadcast soccer tracking with event data and in-venue-supervised spatiotemporal axial attention to reconstruct player and ball trajectories.
Problem
Reconstructing noisy, heavily occluded broadcast tracking trajectories of soccer players and the ball; prior methods use only short (≤10 s) unimodal trajectory context.
Approach
Applies spatiotemporal axial attention to long (60 s) multi-agent trajectory windows and fuses them with event data via per-agent temporal cross-attention, so the ball can be inferred from event context rather than player motion alone.
Key results
- Event2Tracking (60s) achieves 3.22 m player mADE versus 4.25 m for STT at 10s context, and improves on STT by 32-36% for ball mADE across context lengths.
- Player mADE for Event2Tracking improves 22% from 10s to 60s context, versus 15% for the STT baseline.
- Ball mADE for Event2Tracking at 60s is 3.51 m compared with 5.46 m for the next best model (STT at 60s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Event2Tracking instead of STT | Soccer trajectory reconstruction with 60s context · 60s context | mADE (players): 3.62 -> 3.22 | Helped |
| Use Event2Tracking instead of STT | Soccer trajectory reconstruction with 60s context · 60s context | mADE (ball): 5.46 -> 3.51 | Helped |
| Increase context from 10s to 60s (ablation) | Event2Tracking · Soccer trajectory reconstruction | mADE (players): 4.13 -> 3.22 | Helped |
| Increase context from 10s to 60s (ablation) | Event2Tracking · Soccer trajectory reconstruction | mADE (ball): 4.24 -> 3.51 | Helped |
| Increase context from 10s to 60s (ablation) | Spatiotemporal Transformer (STT) · Soccer trajectory reconstruction | mADE (players): 4.25 -> 3.62 | Helped |
| Increase context from 10s to 60s (ablation) | Spatiotemporal Transformer (STT) · Soccer trajectory reconstruction | mADE (ball): 6.23 -> 5.46 | Helped |
Surprising findings
- The Graph Imputer's performance plateaus at 5% improvement from 10s to 30s and degrades for longer segments.
Limitations
- Trajectories longer than 60 s were not tested due to computational constraints.
- Relies on proprietary in-venue tracking and human-annotated event data that may not be widely available.
- Only deterministic baselines were compared; the model is deterministic and does not generate diverse trajectories.
Experimental setup
Datasets: proprietary soccer event data broadcast tracking data in-venue tracking data (700 train / 52 eval games)
Benchmarks: in-house soccer broadcast tracking reconstruction (52 games)
Metrics: mean average displacement error (mADE) for players and ball
Baselines: Linear interpolator Independent Transformer Graph Imputer Spatiotemporal Transformer
Builds on: spatiotemporal axial attention Spatiotemporal Transformer Graph Imputer Wayformer
Closest papers by meaning
- Sports-Traj: A Unified Trajectory Generation Model for Multi-Agent Movement in Sports (ICLR 2025 · similarity 0.90)
- Towards Universal Soccer Video Understanding (CVPR 2025 · similarity 0.88)
- SportR: A Benchmark for Multimodal Large Language Model Reasoning in Sports (ICLR 2026 · similarity 0.87)
- SoccerMaster: A Vision Foundation Model for Soccer Understanding (CVPR 2026 · similarity 0.87)
- Unified Uncertainty-Aware Diffusion for Multi-Agent Trajectory Modeling (CVPR 2025 · similarity 0.86)
- Spot The Ball: A Benchmark for Visual Social Inference (CVPR 2026 · similarity 0.84)
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.