Research map / Detection and segmentation / CNN features
Generative Map Priors for Collaborative BEV Semantic Segmentation
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 CoGMP, a collaborative BEV semantic segmentation framework that compresses features as discrete map-element codebook indices (EFFC) and fuses multi-agent features with a conditional latent diffusion model guided by map structure priors (SGFF).
Problem
Collaborative perception methods are designed for sparse object features and do not handle the dense BEV features of collaborative BEV semantic segmentation, where compression loses information and multi-agent fusion produces conflicting semantics.
Approach
Use a VQ-VAE codebook of map element priors so senders transmit discrete element indices (with learned critical-element selection and masked recovery), and fuse at the receiver by conditioning a latent diffusion U-Net on other agents' element features, initialized from the ego's own features and early-stopped by an uncertainty-adaptive scheduler.
Key results
- CoGMP reaches 66.19/52.92 Road/Lane IoU on OPV2V, a +6.89/+7.64 improvement over prior SOTA when transmitting all intermediate features.
- Achieves a 32-fold communication volume reduction at similar performance, and gains of 3.03/4.33 Road/Lane IoU at 1/512 of baseline bandwidth.
- UATS reduces the average number of SGFF iterations by 4.76 (from 10 to 5.24) with negligible IoU change.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CoGMP instead of previous methods | Collaborative BEV segmentation · OPV2V | Road IoU: +6.89 Road IoU | Helped |
| use CoGMP instead of previous methods | Collaborative BEV segmentation · OPV2V | Lane IoU: +7.64 Lane IoU | Helped |
| use EFFC element-index compression | CoGMP · OPV2V | communication volume: 32-fold reduction | Helped |
| add SGFF to the ego element features (ablation) | CoGMP · OPV2V | Road IoU: 58.76 -> 66.19 | Helped |
| add SGFF to the ego element features (ablation) | CoGMP · OPV2V | Lane IoU: 43.92 -> 52.92 | Helped |
| increase the number of other-agent conditions from 0 to 4 (ablation) | CoGMP SGFF · OPV2V | Road IoU: 60.93 -> 66.19 | Helped |
| increase the number of other-agent conditions from 0 to 4 (ablation) | CoGMP SGFF · OPV2V | Lane IoU: 47.82 -> 52.92 | Helped |
| increase SGFF iterations from 1 to 2 (ablation) | CoGMP SGFF · OPV2V | Road IoU: 59.60 -> 65.28 | Helped |
| add uncertainty-adaptive time scheduling (ablation) | CoGMP SGFF · OPV2V | Road IoU: 66.13 -> 66.19 | Helped |
| add uncertainty-adaptive time scheduling (ablation) | CoGMP SGFF · OPV2V | Lane IoU: 52.75 -> 52.92 | Helped |
| add uncertainty-adaptive time scheduling (ablation) | CoGMP SGFF · OPV2V | SGFF iterations: 10 -> 5.24 | Helped |
| use a generatively trained codebook instead of a task-oriented codebook (ablation) | CoGMP EFFC · OPV2V, communication constraint 28 bytes | Road IoU: 63.03 -> 65.48 | Helped |
Surprising findings
- Even with no other agents (Na=0), the SGFF diffusion refinement alone improves Road/Lane IoU by 2.17/3.9 over the non-fused baseline.
Limitations
- Evaluated only on the simulated OPV2V dataset with camera input
- Diffusion-based fusion remains iterative and costly for perception, mitigated only by UATS and fewer steps
Experimental setup
Datasets: OPV2V
Benchmarks: OPV2V
Metrics: IoU communication volume (log2 bytes)
Baselines: F-Cooper AttFuse V2VNet DiscoNet CoBEVT
Builds on: VQ-VAE CVT ControlNet DDIM CoBEVT
Closest papers by meaning
- DSRC: Learning Density-Insensitive and Semantic-Aware Collaborative Representation Against Corruptions (AAAI 2025 · similarity 0.86)
- An Extensible Framework for Open Heterogeneous Collaborative Perception (ICLR 2024 · similarity 0.85)
- What Makes Good Collaborative Views? Contrastive Mutual Information Maximization for Multi-Agent Perception (AAAI 2024 · similarity 0.85)
- Mutual Learning for SAM Adaptation: A Dual Collaborative Network Framework for Source-Free Domain Transfer (ICML 2025 · similarity 0.85)
- RCP-Bench: Benchmarking Robustness for Collaborative Perception Under Diverse Corruptions (CVPR 2025 · similarity 0.85)
- Privacy-Preserving V2X Collaborative Perception Integrating Unknown Collaborators (AAAI 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.