CL3D: Unsupervised Domain Adaptation for Cross-LiDAR 3D Detection

Domain adaptation for Cross-LiDAR 3D detection is challenging due to the large gap on raw data representation with disparate point densities and arrangements. By exploring domain-invariant geometric characteristics motion patterns, we present an unsupervised domain method that overcomes above difficulties. First, propose Spatial Geometry Alignment module extract similar shape features of same object class align two domains, while eliminating effect distinct distributions. Second, Temporal Motion utilize in sequential frames match domains. Prototypes generated from modules are incorporated into pseudo-label reweighting procedure contribute our effective self-training framework target domain. Extensive experiments show achieves state-of-the-art performance cross-device datasets, especially datasets gaps captured by mechanical scanning LiDARs solid-state various scenes. Project homepage at https://github.com/4DVLab/CL3D.git.