RCP-Bench: Benchmarking Robustness for Collaborative Perception Under Diverse Corruptions

Introduction

Collaborative perception, or Vehicle-to-Vehicle (V2V) communication, enables Connected Autonomous Vehicles (CAVs) to share sensor data, effectively extending their perception range and eliminating blind spots. While promising, current state-of-the-art models are often evaluated under idealized conditions. In real-world deployments, sensors are subject to adverse weather, hardware degradation, and transmission delays. This paper introduces RCP-Bench, the first comprehensive benchmark designed to evaluate the robustness of collaborative perception models against diverse sensor corruptions and temporal misalignments.

The RCP-Bench Framework

To bridge the "robustness gap," the authors established a rigorous testing environment encompassing 14 types of camera-based corruptions. These are categorized into External Weather (e.g., fog, snow, extreme lighting), Camera Interior (e.g., sensor noise, motion blur, and total sensor "crash"), and Temporal Misalignment (simulating data sharing delays).

Expand

Source, p. 3

The benchmark evaluates models across three critical interference scenarios:

1. Global Interference: All vehicles in the network experience corruptions.
2. Ego Interference: Only the primary (ego) vehicle's sensors are corrupted.
3. CAV Interference: Only the collaborating vehicles provide corrupted data.

Robustness Analysis and Key Insights

Evaluation on corrupted versions of major datasets (OPV2V-C, V2XSet-C, and DAIR-V2X-C) revealed several critical vulnerabilities in existing collaborative frameworks.

The "Collaborative Disruption" Risk

A significant finding is that corrupted collaborators can actively degrade the ego vehicle's performance. Many current models lack the mechanism to "distrust" noisy shared data, leading to the propagation of errors throughout the system. As shown in the performance analysis, when collaborators provide corrupted data (CAV Interference), the ego vehicle's performance can drop significantly compared to the clean baseline.

Expand

Source, p. 1

Fusion Method Resilience

The study found that Late Fusion (sharing final object detections) is inherently more resilient than Intermediate Fusion (sharing neural network features). In intermediate fusion, a noisy feature map can easily corrupt the shared latent space, whereas late fusion isolates errors to individual bounding boxes. Furthermore, the results indicate that model complexity does not equate to robustness; simpler "Max Fusion" baselines often outperformed sophisticated attention-based models under high corruption levels.

Proposed Training Strategies: RCP-Drop and RCP-Mix

To mitigate these vulnerabilities, the authors propose two novel training strategies:

• RCP-Drop: A dynamic dropout strategy that randomly discards collaborator features during training. This encourages the ego vehicle to maintain strong independent perception while treating collaborative data as an enhancement rather than a dependency.
• RCP-Mix: This strategy mixes feature statistics (mean and standard deviation) between the ego vehicle and its neighbors, improving the model's ability to generalize across different corruption "styles."

These methods yielded substantial improvements. For instance, the CoBEVT model's corrupted Average Precision ($AP_{cor}$) doubled from 15.91 to 31.83 when trained with RCP-Mix.

Conclusion

RCP-Bench highlights the fragility of current collaborative perception systems in the face of sensor noise and environmental challenges. While strategies like RCP-Drop and RCP-Mix offer significant gains, future work must address cross-modal robustness—such as using LiDAR to compensate for camera failures—and develop more sophisticated data-filtering mechanisms to ensure safety in diverse real-world conditions.