SADA-3D: Structure-Aware Unsupervised Domain Adaptation Segmentation of 3D Point Clouds

Abstract

Domain adaptive LiDAR point cloud segmentation aims to develop an effective target segmentation model using labeled source data and unlabeled target data. Existing domain adaptation methods for segmentation primarily focus on global feature alignment, often neglecting critical structural cues, especially for structure-sensitive categories such as vehicles and pedestrians. This oversight leads to degraded boundary delineation and compromised object structural coherence. To address these challenges, we propose SADA-3D, a simple yet effective structure-aware framework for cross-domain 3D semantic segmentation, with a particular focus on improving segmentation of structure-sensitive categories. First, we introduce a detachable structure-guided feature alignment (SGFA) module to help the backbone network learn structural features that are aligned across both domains. The core idea of SGFA is to leverage geometric priors from auxiliary 3D object detection, which relies on 3D bounding box annotations in the source domain, to explicitly model feature alignment structural patterns while performing category-aware feature alignment in bird's-eye-view (BEV) space. Second, we design a point-level feature alignment(PLFA) module to enable fine-grained domain adaptation. This module leverages the voxel features output by the decoder, in combination with the candidate bounding boxes generated by the SGFA module, to achieve cross-domain feature alignment for foreground points via a domain discriminator. Additionally, we incorporate a BEV auxiliary loss to further enhance the model's overall domain generalization ability. Our method has been extensively evaluated on multiple public datasets, demonstrating state-of-the-art performance, with especially significant improvements for structure-sensitive categories such as vehicles and pedestrians.