Research map / Label-efficient and robust learning / Recognition and long-tail
RCP-Bench: Benchmarking Robustness for Collaborative Perception Under Diverse Corruptions
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.
Introduces RCP-Bench, a benchmark of 14 camera corruptions across six collaborative scenarios on OPV2V, V2XSet and DAIR-V2X, plus two training strategies (RCP-Drop, RCP-Mix) to improve collaborative perception robustness.
Problem
Collaborative perception models are typically evaluated under ideal conditions, so their robustness to real-world camera corruptions (weather, sensor failures, temporal misalignment) in multi-vehicle settings is largely unknown.
Approach
Builds corrupted versions of three V2X perception datasets with 14 camera corruption types at 5 severities and evaluates Global, Ego and CAV interference scenarios; proposes RCP-Drop (randomly dropping CAV features during training with a decaying drop probability) and RCP-Mix (mixing feature mean/std between ego and CAV features) as simple robustness boosters.
Key results
- Across 10 collaborative models on OPV2V-C, mRCE generally lies between 55% and 70%, indicating suboptimal robustness under corruption.
- Adding BN plus RCP-Mix to CoBEVT raises mean APcor from 15.91 to 31.83 and APclean from 40.49 to 48.25.
- Adding BN plus RCP-Drop to AttFuse raises mean APcor from 15.99 to 27.56 and reduces mNegC from 121.59 to 115.64.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Late fusion instead of intermediate feature fusion | Collaborative perception models · OPV2V-C Global Interference | APcor@0.5: 26.62 vs. 22.21 | Helped |
| Add RCP-Mix | CoBEVT + BN · OPV2V-C | APcor@0.5: 27.87 -> 31.83 | Helped |
| Add RCP-Drop | CoBEVT + BN · OPV2V-C | APcor@0.5: 27.87 -> 30.68 | Helped |
| Add RCP-Mix | AttFuse + BN · OPV2V-C | APcor@0.5: 22.35 -> 28.85 | Helped |
| Add RCP-Drop | AttFuse + BN · OPV2V-C | APcor@0.5: 22.35 -> 27.56 | Helped |
| Replace ResNet101 with EfficientNet-b0 (ablation) | Max · OPV2V-C | AP@0.5: 35.12 -> 45.87 | Helped |
| Replace ResNet101 with EfficientNet-b0 (ablation) | V2X-ViT · OPV2V-C | AP@0.5: 46.21 -> 58.61 | Helped |
| Add multiscale fusion (ablation) | Max · OPV2V-C | APcor@0.5: 18.05 -> 33.22 | Helped |
| Add multiscale fusion (ablation) | AttFuse · OPV2V-C | APcor@0.5: 14.93 -> 24.08 | Helped |
| Increase the number of collaborating vehicles (ablation) | Collaborative perception models · OPV2V-C | APcor | Helped |
| Increase the number of cameras (ablation) | Collaborative perception models · OPV2V-C | APcor | Helped |
Surprising findings
- Simple Max fusion and Late fusion outperform several more complex intermediate fusion models under interference.
- Under New Scenes with Corruption, V2X-ViT's APclean drops 47.8% when moving to unseen scenes, and corruption adds further degradation.
Limitations
- Only camera-based corruptions are studied; lidar corruptions are excluded due to mutual interference among CAVs.
- RCP-Drop and RCP-Mix are evaluated only on CoBEVT and AttFuse with BN adaptation.
- Corruptions are simulated on validation sets rather than collected from real sensor failures.
Experimental setup
Datasets: OPV2V V2XSet DAIR-V2X
Benchmarks: OPV2V-C V2XSet-C DAIR-V2X-C
Metrics: AP@0.5 AP@0.3 AP@0.7 APcor mRCE mPosC mNegC
Baselines: AttFuse F-Cooper V2X-ViT DiscoNet V2VNet CoBEVT BM2CP Max fusion Late fusion NoFusion
Builds on: MixStyle Dropout ImageNet-C Stats Adapt (online BN)
Closest papers by meaning
- CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization (CVPR 2025 · similarity 0.89)
- ERMVP: Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments (CVPR 2024 · similarity 0.88)
- DSRC: Learning Density-Insensitive and Semantic-Aware Collaborative Representation Against Corruptions (AAAI 2025 · similarity 0.88)
- RCDN: Towards Robust Camera-Insensitivity Collaborative Perception via Dynamic Feature-based 3D Neural Modeling (NeurIPS 2024 · similarity 0.87)
- STAMP: Scalable Task- And Model-agnostic Collaborative Perception (ICLR 2025 · similarity 0.86)
- AirCopBench: A Benchmark for Multi-drone Collaborative Embodied Perception and Reasoning (AAAI 2026 · 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.