Research map / Video / Video segmentation and tracking
BEVSync: Asynchronous Data Alignment for Camera-based Vehicle-Infrastructure Cooperative Perception Under Uncertain Delays
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 BEVSync, a camera-based V2I cooperative perception system that uses an attention-based extractor-compensator to predict future infrastructure BEV features and align asynchronous frames under uncertain transmission delays and clock differences.
Problem
Camera-based vehicle-infrastructure cooperative 3D perception suffers from time asynchrony between ego-vehicle and roadside image frames caused by clock differences and variable processing/transmission delays, which degrades fusion accuracy.
Approach
Infrastructure extracts motion information from three historical BEV frames via deformable attention and sends it with timestamps; the ego-vehicle uses a time-conditioned compensator to predict the synchronous infrastructure BEV feature before spatial warping and fusion.
Key results
- Under 300 ms transmission delay, BEVSync reaches 28.33 mAP@3D (IoU0.3) versus 26.59 for FFNet and 19.05 for VIMI.
- With clock differences uniformly sampled in [-50, 50] ms and 300 ms delay, BEVSync keeps 28.25 mAP@3D (IoU0.3), close to its synchronized result.
- Window size ablation: mAP@3D (IoU0.3) rises from 26.19 at 2 frames to 28.33 at 3 frames and 29.08 at 5 frames.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BEVSync instead of FFNet | Camera-based V2I cooperative perception · DAIR-V2X test, 100 ms delay, 0 ms clock difference | mAP@3D IoU0.3: 30.18 -> 31.59 | Helped |
| use BEVSync instead of FFNet | Camera-based V2I cooperative perception · DAIR-V2X test, 300 ms delay, 0 ms clock difference | mAP@3D IoU0.3: 26.59 -> 28.33 | Helped |
| use BEVSync instead of FFNet | Camera-based V2I cooperative perception · DAIR-V2X test, 300 ms delay, [-50, 50] ms clock difference | mAP@3D IoU0.3: 26.20 -> 28.25 | Helped |
| add the extractor (ablation) | BEVSync · DAIR-V2X test, 0 ms delay | mAP@3D IoU0.3: 31.72 -> 33.40 | Helped |
| add the extractor (ablation) | BEVSync · DAIR-V2X test, 300 ms delay | mAP@3D IoU0.3: 18.91 -> 28.33 | Helped |
| add the compensator (ablation) | BEVSync · DAIR-V2X test, 300 ms delay | mAP@3D IoU0.3: 19.29 -> 28.33 | Helped |
| use window size 2 instead of 3 (ablation) | BEVSync · DAIR-V2X test, 300 ms delay | mAP@3D IoU0.3: 28.33 -> 26.19 | Hurt |
| increase window size from 3 to 4 (ablation) | BEVSync · DAIR-V2X test, 300 ms delay | mAP@3D IoU0.3: 28.33 -> 28.76 | Helped |
| increase window size from 4 to 5 (ablation) | BEVSync · DAIR-V2X test, 300 ms delay | mAP@3D IoU0.3: 28.76 -> 29.08 | Helped |
Surprising findings
- Without the extractor, BEVSync still improves over no-compensation only modestly at 0 ms, but the extractor also gives a small gain even in synchronous settings, attributed to occluded objects visible in history frames.
- With large delays above 300 ms, synchronous intermediate fusion (VIMI) can perform worse than the vehicle-only non-fusion baseline.
Limitations
- Evaluated only on the camera-based part of DAIR-V2X with simulated delays by using past infrastructure frames.
- Clock differences modeled only within an assumed NTP range of [-50, 50] ms; larger errors are not tested.
- Compensation relies on historical motion and performs worse for non-linear or unpredictable dynamics, and results with window size 2 degrade to linear prediction.
Experimental setup
Datasets: DAIR-V2X
Benchmarks: DAIR-V2X
Metrics: mAP@3D IoU0.3 mAP@3D IoU0.5 mAP@BEV IoU0.3 mAP@BEV IoU0.5
Baselines: Vehicle-only Early Fusion VIMI Late + Kalman (VIPS) FFNet
Builds on: BEVHeight BEVDepth BEVFormer Deformable DETR VIMI FFNet
Closest papers by meaning
- TraF-Align: Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception (CVPR 2025 · similarity 0.89)
- V2X-Radar: A Multi-modal Dataset with 4D Radar for Cooperative Perception (NeurIPS 2025 · similarity 0.87)
- UrbanIng-V2X: A Large-Scale Multi-Vehicle, Multi-Infrastructure Dataset Across Multiple Intersections for Cooperative Perception (NeurIPS 2025 · similarity 0.86)
- TUMTraf V2X Cooperative Perception Dataset (CVPR 2024 · similarity 0.85)
- AGC-Drive: A Large-Scale Dataset for Real-World Aerial-Ground Collaboration in Driving Scenarios (NeurIPS 2025 · similarity 0.85)
- SparseAlign: a Fully Sparse Framework for Cooperative Object Detection (CVPR 2025 · similarity 0.85)
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.