Research map / Multi-agent systems and games / Human-AI and goal-conditioned
TraF-Align: Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception
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 TraF-Align, a trajectory-aware feature alignment framework that predicts feature-level object trajectory fields and uses offset-guided attention along them to compensate for inter-agent latency in cooperative LiDAR perception.
Problem
Inter-agent communication latency causes spatial and semantic misalignment of features in asynchronous multi-agent cooperative 3D object detection.
Approach
Predicts a trajectory field (position and orientation heatmaps) of objects up to the ego's current time, generates attention sampling offsets along each trajectory, and applies attention from current-time queries to historical features on the same trajectory to reconstruct current features and align semantics across agents, trained end-to-end with field and offset losses.
Key results
- On V2V4Real at 0 ms, TraF-Align reaches 74.28/44.14 AP50/70 vs. 70.74/38.45 for ERMVP (the best prior feature-level method).
- At 400 ms latency on V2V4Real, TraF-Align drops only 4.87/5.65 AP points from synchronous performance, versus 13.48/9.14 for ERMVP.
- On DAIR-V2X-Seq at 400 ms, TraF-Align achieves 71.22/52.01 AP50/70, compared with 59.55/43.93 for ERMVP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TraF-Align instead of ERMVP | cooperative perception · V2V4Real val, 0 ms | AP50/AP70: 70.74/38.45 -> 74.28/44.14 | Helped |
| use TraF-Align instead of ERMVP | cooperative perception · DAIR-V2X-Seq val, 400 ms | AP50/AP70: 59.55/43.93 -> 71.22/52.01 | Helped |
| add field predictor (ablation) | TraF-Align · DAIR-V2X-Seq, infrastructure latency 0 ms | AP50/AP70: 75.35/54.72 -> 76.31/57.02 | Helped |
| add offset generator (ablation) | TraF-Align with field predictor · DAIR-V2X-Seq, infrastructure latency 0 ms | AP50/AP70: 76.31/57.02 -> 77.11/58.63 | Helped |
| add attention layers (ablation) | TraF-Align with field predictor and offset generator · DAIR-V2X-Seq, infrastructure latency 0 ms | AP50/AP70: 77.11/58.63 -> 77.83/59.72 | Helped |
| add field loss (ablation) | TraF-Align without offset loss · DAIR-V2X-Seq, infrastructure latency 0 ms | AP50/AP70: 76.43/57.74 -> 77.12/58.36 | Helped |
| add offset loss (ablation) | TraF-Align without field loss · DAIR-V2X-Seq, infrastructure latency 0 ms | AP50/AP70: 76.43/57.74 -> 77.15/58.70 | Helped |
| add field loss and offset loss (ablation) | TraF-Align · DAIR-V2X-Seq, infrastructure latency 0 ms | AP50/AP70: 76.43/57.74 -> 77.83/59.72 | Helped |
| increase cooperative vehicle frames from 1 to 4 (ablation) | TraF-Align with 2 ego frames · V2V4Real, cooperative vehicle latency 400 ms | AP50/AP70: 64.28/32.28 -> 69.41/38.49 | Helped |
| increase ego frames from 2 to 4 (ablation) | TraF-Align with 4 cooperative vehicle frames · V2V4Real, cooperative vehicle latency 400 ms | AP50/AP70: 69.41/38.49 -> 69.08/37.08 | Hurt |
Surprising findings
- Applying latency to the vehicle agent yields lower AP than applying it to the infrastructure agent, attributed to differing object visibility distributions between agents.
Limitations
- Trained mainly on regular latencies (multiples of 100 ms), not realistic latency distributions.
- Detection precision under delay can still be improved; multi-frame inputs introduce motion blur.
- Relies on accurate inter-frame and inter-agent pose alignment; pose errors degrade performance.
Experimental setup
Datasets: V2V4Real DAIR-V2X-Seq
Benchmarks: V2V4Real DAIR-V2X-Seq
Metrics: AP50 AP70
Baselines: Single Agent AttFuse F-Cooper V2VNet Where2comm V2X-ViT MRCNet CoBEVT ERMVP
Builds on: PointPillars PillarNext Transformer Sinkhorn algorithm CoBEVFlow
Closest papers by meaning
- BEVSync: Asynchronous Data Alignment for Camera-based Vehicle-Infrastructure Cooperative Perception Under Uncertain Delays (AAAI 2025 · similarity 0.89)
- Asynchrony-Robust Collaborative Perception via Bird's Eye View Flow (NeurIPS 2023 · similarity 0.87)
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.84)
- STAMP: Scalable Task- And Model-agnostic Collaborative Perception (ICLR 2025 · similarity 0.84)
- One is Plenty: A Polymorphic Feature Interpreter for Immutable Heterogeneous Collaborative Perception (CVPR 2025 · similarity 0.84)
- LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (NeurIPS 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.