Abstract
As two fundamental representation modalities of 3D objects, 2D multi-view images and 3D point clouds reflect shape information from different aspects of visual appearances and geometric structures. Unlike deep learning-based 2D multi-view image modeling, which demonstrates leading performances in various 3D shape analysis tasks, 3D point cloud-based geometric modeling still suffers from insufficient learning capacity. In this paper, we innovatively construct a unified cross-modal knowledge transfer framework, which distills discriminative visual descriptors of 2D images into geometric descriptors of 3D point clouds. Technically, under a classic teacher-student learning paradigm, we propose multi-view vision-to-geometry distillation, consisting of a deep 2D image encoder as teacher and a deep 3D point cloud encoder as student. To achieve heterogeneous feature alignment, we further propose visibility-aware feature projection, through which per-point embeddings can be aggregated into multi-view geometric descriptors. Extensive experiments on 3D shape classification, part segmentation, and unsupervised learning validate the superiority of our method. We will make the code and data publicly available.
Abstract (translated)
URL
https://arxiv.org/abs/2207.03128
