RefRec: Pseudo-labels Refinement via Shape Reconstruction for Unsupervised 3D Domain Adaptation

Unsupervised Domain Adaptation (UDA) for point cloud classification is an\nemerging research problem with relevant practical motivations. Reliance on\nmulti-task learning to align features across domains has been the standard way\nto tackle it. In this paper, we take a different path and propose RefRec, the\nfirst approach to investigate pseudo-labels and self-training in UDA for point\nclouds. We present two main innovations to make self-training effective on 3D\ndata: i) refinement of noisy pseudo-labels by matching shape descriptors that\nare learned by the unsupervised task of shape reconstruction on both domains;\nii) a novel self-training protocol that learns domain-specific decision\nboundaries and reduces the negative impact of mislabelled target samples and\nin-domain intra-class variability. RefRec sets the new state of the art in both\nstandard benchmarks used to test UDA for point cloud classification, showcasing\nthe effectiveness of self-training for this important problem.\n