Toward Robust Collaborative 3-D Object Detection via Two-Stage Cross-Domain Knowledge Transfer in Autonomous Driving

Abstract

Recent research has demonstrated that Vehicle-to-Everything (V2X) collaborative 3D object detection can provide sufficient scene understanding to support autonomous driving. However, these methods suffer from severe performance degradation when generalized to diverse environments and deployment conditions. Existing 3D adaptive methods tend to transfer from a homogeneous source domain knowledge to the target domain. Yet, the scarcity of well-annotated collaborative datasets and the Simulation-to-Reality (Sim2Real) domain gap limit their effectiveness. To this end, we propose a novel two-stage unsupervised domain adaptation method, S2C-UDA, for multi-agent 3D object detection driven by one single-agent source. Concretely, we first design an agent-wise Joint Domain Adversarial Pre-training network (JDA-PT) to align each target agent with the source domain, thus efficiently releasing pseudo labels. Then comes the customized Multi-level Response Self-training network (MLR-ST) to enhance intra-domain cross-agent adaptation by sequentially integrating feature-, instance-, and result-level responses. This hierarchical framework systematically aligns inter-agent feature distributions while simultaneously mitigating instance discrepancies, thereby facilitating the iteration of reliable detection results. Extensive evaluations on challenging real-world datasets reveal that S2C-UDA achieves superior adaptation performance for cross-domain collaborative 3D object detection in outdoor traffic scenarios.