nuScenes (Cross-City UDA)

ConDA: Unsupervised Domain Adaptation for LiDAR Segmentation via Regularized Domain Concatenation

ConDA is presented, a concatenation-based domain adaptation framework for LiDAR segmentation that constructs an intermediate domain consisting of fine-grained interchange signals from both source and target domains without destabilizing the semantic coherency of objects and background around the ego-vehicle and significantly outperforms prior arts in mitigating domain gaps.