CUDA-X: Unsupervised Domain-Adaptive Vehicle-to-Everything Collaboration via Knowledge Transfer and Alignment

Recently emerged vehicle-to-everything (V2X) perception has revealed great potential to overcome the limitation of single-vehicle intelligence aided by vigorous interaction among on-road agents, while prior endeavors are practically developed on parameter-specific simulation or configuration-dynamic real-world setting, overlooking the transferability across various scenarios. In this article, we propose unsupervised domain-adaptive vehicle-to-everything collaboration framework dubbed CUDA-X, which is built on top of a de facto collective model with key-point information exchange and instance adaptation. Specifically, collaborative knowledge transfer (CKT) is responsible for domain-agnostic feature reconstruction from nearby car or infrastructure by spatial–channel pooling operation in an elementwise manner. To promote the candidate alignment, a brand-new bin-based location correction (BLC) provides an auxiliary supervision for cross-dataset box refinement via residual coordinate encoding (RCE), and category-aware pooling alignment (CPA) is further designed for pulling the category-specific instance closer between source and target samples. We benchmark CUDA-X against the counterparts on four prevalent cooperative perception datasets, i.e., OPV2V, V2X-Sim, V2V4Real, and DAIR-V2X: it establishes the new state-of-the-art vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) performances regardless of simulation or reality. We expect that this appealing attempt would provide an in-depth insight into domain generalization in the context of multiagent perception, and the code is publicly available soon.