Research map / Multi-agent systems and games / Human-AI and goal-conditioned
SparScene: Efficient Traffic Scene Representation via Sparse Graph Learning for Large-Scale Trajectory Generation
Xiaoyu Mo, Jintian Ge, Zifan Wang, Chen Lv, Karl Henrik Johansson
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 TiL module to L2A-only variant (ablation) | SparScene (AL+G2) · WOMD val (20% train) | minFDE@8s (m): 3.28 -> 2.89 | Improved |
| add A2A interaction module (ablation) | SparScene (AL+G12) · WOMD val (20% train) | minFDE@8s (m): 2.89 -> 2.73 | Improved |
| enlarge A2L radius from 5 to 10 (ablation) | SparScene · WOMD val (20% train) | minFDE@8s (m) | Improved |
| enlarge A2L radius from 10 to 30 and 50 (ablation) | SparScene · WOMD val (20% train) | minFDE@8s (m) | No clear effect |
| use dense three-hop omnidirectional search (OOO) instead of OFF (ablation) | SparScene · WOMD val (20% train), r=10 | minFDE@8s (m): 2.77 -> 2.78 | No clear effect |
| use OFF search instead of purely forward (F, FFF) or one-step omnidirectional (O) (ablation) | SparScene · WOMD val (20% train), r=10 | minFDE@8s (m): 2.77 | Improved |
| scene-level single-pass generation vs per-agent MTR++ as agents grow from 8 to 32 | SparScene · WOMD val (scaling, focal agents) | latency (ms): 77 -> 118 (MTR++) vs 5 -> 5 (SparScene) | Improved |
| sparse lightweight scene graph vs MTR++ | SparScene · WOMD test | latency (ms): 118 ms -> 5 ms; 86.6 M -> 3.2 M params | Improved |
| sparse scene graph vs MTR++ on accuracy | SparScene · WOMD test | MR: 0.1298 -> 0.1559 | Hurt |
| share encoded context across agents (symmetric encoder, MTR to MTR+) | MTR · WOMD val (scaling, focal agents) | latency (ms): 193 ms -> 98 ms; 15.6 GB -> 4.7 GB | Improved |
| batch scenes into larger graph up to 5,620 agents | SparScene · Regional scaling (batched WOMD scenes) | latency (ms): 5 -> 54 | Improved |
Closest papers by meaning
- VectorNet: Encoding HD Maps and Agent Dynamics From Vectorized Representation (CVPR 2020 · similarity 0.88)
- HDGT: Heterogeneous Driving Graph Transformer for Multi-Agent Trajectory Prediction via Scene Encoding (TPAMI 2023 · similarity 0.88)
- SEPT: Towards Efficient Scene Representation Learning for Motion Prediction (ICLR 2024 · similarity 0.87)
- Trajeglish: Traffic Modeling as Next-Token Prediction (ICLR 2024 · similarity 0.86)
- T2SG: Traffic Topology Scene Graph for Topology Reasoning in Autonomous Driving (CVPR 2025 · similarity 0.86)
- SceneDiffuser++: City-Scale Traffic Simulation via a Generative World Model (CVPR 2025 · similarity 0.86)
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.
