UniDA3D: Unified Domain Adaptive 3D Semantic Segmentation Pipeline

원문 초록

State-of-the-art 3D semantic segmentation models are trained on off-the-shelf public benchmarks, but they will inevitably face the challenge of recognition accuracy drop when these well-trained models are deployed to a new domain. In this paper, we introduce a Unified Domain Adaptive 3D semantic segmentation pipeline (UniDA3D) to enhance the weak generalization ability, and bridge the point distribution gap between domains. Different from previous studies that only focus on a single adaptation task, UniDA3D can tackle several adaptation tasks in 3D segmentation field, by designing a unified source-and-target active sampling strategy, which selects a maximally-informative subset from both source and target domains for effective model adaptation. Besides, benefiting from the rise of multi-modal 2D-3D datasets, UniDA3D investigates the possibility of achieving a multi-modal sampling strategy, by developing a cross-modality feature interaction module that can extract a representative pair of image and point features to achieve a bi-directional image-point feature interaction for safe model adaptation. Experimentally, UniDA3D is verified to be effective in many adaptation tasks including: 1) unsupervised domain adaptation, 2) unsupervised few-shot domain adaptation; 3) active domain adaptation. Their results demonstrate that, by easily coupling UniDA3D with off-the-shelf 3D segmentation baselines, domain generalization ability of these baselines can be enhanced.