Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Sports-Traj: A Unified Trajectory Generation Model for Multi-Agent Movement in Sports
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 UniTraj, a CVAE-based masked-trajectory generator with Ghost Spatial Masking in a Transformer and Bidirectional Temporal Mamba with a Bidirectional Temporal Scaled module, unifying prediction, imputation, and spatial-temporal recovery, and releases three sports datasets (Basketball-U, Football-U, Soccer-U).
Problem
Multi-agent trajectory modeling is usually handled per task (prediction, imputation, spatial-temporal recovery), and existing sports and pedestrian datasets do not cover structured multi-agent interactions with diverse missing patterns.
Approach
Treat any trajectory input as a masked sequence and generate the complete trajectory with a CVAE; a ghost masking token summarizes the missing pattern across agents in a Transformer, while a bidirectional Mamba with a decay-based temporal scaling module models time gaps in both directions.
Key results
- On Basketball-U, UniTraj reaches minADE20 4.77 vs 5.81 for the best baseline GC-VRNN (about 17.9% lower).
- On Football-U, minADE20 3.55 vs 4.95 for GC-VRNN (about 28.3% lower); on Soccer-U 94.59 vs 105.87 (about 10.7% lower).
- On ETH-UCY, average minADE20/minFDE20 is 0.23/0.36, slightly behind MemoNet and EqMotion (0.21/0.35), ahead of FlowChain (0.29/0.52).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UniTraj instead of GC-VRNN | sports trajectory generation · Basketball-U | minADE20: 5.81 -> 4.77 | Helped |
| use UniTraj instead of GC-VRNN | sports trajectory generation · Football-U | minADE20: 4.95 -> 3.55 | Helped |
| use UniTraj instead of GC-VRNN | sports trajectory generation · Soccer-U | minADE20: 105.87 -> 94.59 | Helped |
| add Ghost Spatial Masking (ablation) | UniTraj · Basketball-U, Football-U, Soccer-U | minADE20: 4.86 -> 4.77; 3.92 -> 3.55; 119.43 -> 94.59 | Helped |
| add Bidirectional Temporal Scaled module (ablation) | UniTraj with bidirectional Mamba · Basketball-U, Football-U, Soccer-U | minADE20: 4.86 -> 4.77; 3.60 -> 3.55; 105.47 -> 94.59 | Helped |
| replace bidirectional Mamba with forward unidirectional Mamba using BTS (ablation) | UniTraj · Basketball-U, Football-U, Soccer-U | minADE20: 4.77 -> 5.86; 3.55 -> 4.09; 94.59 -> 106.22 | Hurt |
| replace max-pooling with mean-pooling for ghost embedding (ablation) | UniTraj · Basketball-U, Football-U, Soccer-U | minADE20: 4.77 -> 4.80; 3.55 -> 3.64; 94.59 -> 100.84 | Mixed |
| replace max-pooling with sum-pooling for ghost embedding (ablation) | UniTraj · Basketball-U, Football-U, Soccer-U | minADE20: 4.77 -> 4.79; 3.55 -> 3.56; 94.59 -> 102.99 | Mixed |
| replace temporal Mamba with LSTM (ablation) | UniTraj · Basketball-U | minADE20: 5.32 vs 4.77 | Hurt |
| replace temporal Mamba with VRNN (ablation) | UniTraj · Basketball-U | minADE20: 5.29 vs 4.77 | Hurt |
| replace temporal Mamba with Transformer (ablation) | UniTraj · Basketball-U | minADE20: 4.99 vs 4.77 | Hurt |
| set Mamba block depth to L=5 rather than L=4 (ablation) | UniTraj · Basketball-U | minADE20: 4.77 -> 4.81 | Hurt |
Surprising findings
- Linear Fit matches ground truth closely on Step, Path-L and Path-D metrics, showing these length-based metrics fail to discriminate generation quality.
- On Soccer-U, the w/o BTS variant (minADE20 105.47) outperforms the full model on Soccer-U only in the ablation table's ordering, indicating component contributions vary across datasets.
Limitations
- Decoder is a simple MLP, which limits generation quality.
- Fixed number of agents in the sports datasets; varying agent counts not handled.
- Results for the INAM baseline are reproductions since official code was not released.
Experimental setup
Datasets: NBA tracking data (Stats Perform) NFL Big Data Bowl / Next Gen Stats tracking SoccerTrack ETH-UCY SDD Traffic-Guangzhou
Benchmarks: Basketball-U Football-U Soccer-U ETH-UCY SDD Traffic-Guangzhou
Metrics: minADE20 OOB Step Path-L Path-D minFDE20 RMSE MAE
Baselines: Mean Median Linear Fit LSTM Transformer MAT NAOMI INAM SSSD GC-VRNN
Builds on: Mamba Transformer CVAE NAOMI SSSD
Closest papers by meaning
- Event2Tracking: Reconstructing Multi-Agent Soccer Trajectories Using Long-Term Multimodal Context (AAAI 2025 · similarity 0.90)
- Unified Uncertainty-Aware Diffusion for Multi-Agent Trajectory Modeling (CVPR 2025 · similarity 0.87)
- Social-Transmotion: Promptable Human Trajectory Prediction (ICLR 2024 · similarity 0.85)
- Samba: Synchronized Set-of-Sequences Modeling for Multiple Object Tracking (ICLR 2025 · similarity 0.84)
- SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model (CVPR 2024 · similarity 0.84)
- FineSports: A Multi-person Hierarchical Sports Video Dataset for Fine-grained Action Understanding (CVPR 2024 · 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.