An Extensible Framework for Open Heterogeneous Collaborative Perception

TL;DR

This work tackles open heterogeneous collaborative perception by introducing HEAL, a framework that preserves a unified feature space for multi-agent collaboration and enables continually emerging agent types to join with ultra-low training costs via backward alignment. HEAL's core components are a Pyramid Fusion-based collaboration base training that builds a robust, multi-scale, foreground-aware shared space, and a backward alignment step that adapts new agents’ front-end encoders to this space without retraining the entire model. The authors also introduce OPV2V-H, a large-scale heterogeneous dataset to benchmark cross-modal collaboration, and demonstrate SOTA performance with substantial reductions in training parameters (up to 91.5%) when integrating multiple new agents, on both OPV2V-H and DAIR-V2X. The approach offers practical benefits for real-world deployment, including privacy preservation and scalable expansion of collaborative perception across diverse sensor configurations and models.

Abstract

Collaborative perception aims to mitigate the limitations of single-agent perception, such as occlusions, by facilitating data exchange among multiple agents. However, most current works consider a homogeneous scenario where all agents use identity sensors and perception models. In reality, heterogeneous agent types may continually emerge and inevitably face a domain gap when collaborating with existing agents. In this paper, we introduce a new open heterogeneous problem: how to accommodate continually emerging new heterogeneous agent types into collaborative perception, while ensuring high perception performance and low integration cost? To address this problem, we propose HEterogeneous ALliance (HEAL), a novel extensible collaborative perception framework. HEAL first establishes a unified feature space with initial agents via a novel multi-scale foreground-aware Pyramid Fusion network. When heterogeneous new agents emerge with previously unseen modalities or models, we align them to the established unified space with an innovative backward alignment. This step only involves individual training on the new agent type, thus presenting extremely low training costs and high extensibility. To enrich agents' data heterogeneity, we bring OPV2V-H, a new large-scale dataset with more diverse sensor types. Extensive experiments on OPV2V-H and DAIR-V2X datasets show that HEAL surpasses SOTA methods in performance while reducing the training parameters by 91.5% when integrating 3 new agent types. We further implement a comprehensive codebase at: https://github.com/yifanlu0227/HEAL

Figures (10)

• Figure 1: (a) homogeneous setting, where agents have identical modality and model. (b) heterogeneous setting, where agents' modalities and models are distinct but pre-determined. (c) Open heterogeneous setting, where new types of agents want to join collaboration with previously unseen modalities or models. (d) HEAL holds the SOTA performance while minimizing the training cost (model parameters here) when integrating a new agent type. The bullseye represents the best.
• Figure 2: Overview of HEAL. (i) We train the initial homogeneous agents (collaboration base) with our novel Pyramid Fusion to establish a unified feature space; (ii) We leverage the well-trained Pyramid Fusion and detection head as the new agents' detection back-end. With the back-end fixed, it pushes the encoder to align its features within the unified feature. This step is performed on the new agent type only, presenting extremely low training costs. (iii) New agents join the collaboration.
• Figure 3: Pyramid Fusion uses multiscale and foreground-aware designs to fuse features and create a robust unified feature space. Foreground estimators produce foreground possibility maps at each BEV position. These foreground maps are then normalized to weights for feature summation. Foreground maps are subject to supervision during training. Blue and green represent different agents.
• Figure 4: Robust Experiment to pose error and compression ratio. Pose noise is set to $\mathcal{N}(0, \sigma_p^2)$ on x, y location and $\mathcal{N}(0, \sigma_r^2)$ on yaw angle.
• Figure 5: Visualization of HEAL's backward alignment from L2 to L1's unified space.