Collaborative perception improves environmental understanding by sharing sensory information among multiple agents, yet real‑world deployment is hindered by heterogeneous sensor modalities and model architectures. Existing two‑stage protocol methods map heterogeneous features into a shared protocol space, but independently trained modality‑specific converters produce pseudo‑protocol distributions that are modality‑biased, causing semantic inconsistency and error accumulation, especially when modality gaps are large.
To tackle this, we introduce CauseCollab, a causal unified and modality‑agnostic network. From a causal perspective it learns representations in the protocol space and employs causal metric learning to disentangle semantic factors $S$ from modality‑specific statistical confounders $C$, with the loss formulation $$\mathcal{L}=\mathcal{L}_{task}+\lambda\,\mathcal{L}_{causal},\quad \mathcal{L}_{causal}=\|\mathbf{E}[S|C]-\mathbf{E}[S]\|^{2}.$$
The architecture uses a context‑guided Unified Converter to ensure cross‑modal semantic consistency, and new modalities can be added by training lightweight adapters with minimal parameters.
Extensive experiments on the OPV2V and DAIR‑V2X benchmarks show that CauseCollab achieves state‑of‑the‑art results, with particularly large gains in scenarios with substantial modality gaps.
Review