GeoDP: Geometry-anchored Dynamic Prompt for Heterogeneous Collaborative Perception

Collaborative perception improves the perception performance of individual agents through information sharing, but feature heterogeneity caused by different sensors and perception models poses significant challenges for practical applications. Existing methods mainly focus on pairwise adapters or naive feature-space alignment, which either scale poorly as model diversity increases or lack essential geometric information and cannot dynamically adapt to different models. To address these issues, we propose GeoDP, a Geometry-anchored Dynamic Prompt framework for heterogeneous collaborative perception. Specifically, we leverage a geometry foundation model to construct a Geometry-anchored Unified Space (GUS), which provides the essential geometric information required for 3D collaborative perception and offers a unified reference feature space for heterogeneous agents. Furthermore, a Dynamic Prompt Adapter (DPA) generates input-conditioned prompts based on the discrepancy between each agent feature and the geometry anchor, scaling flexibly and dynamically to diverse models. Finally, Adaptive Spatial Fusion (ASF) assigns different weights to the adapted features for reliable collaboration. Extensive experiments on OPV2V and DAIR-V2X demonstrate that GeoDP achieves new state-of-the-art performance across both seen and unseen models, outperforming existing methods by a large margin and exhibiting strong robustness. The code and model will be publicly available.