Mutual information-driven self-supervised point cloud pre-training

Weichen Xu, Tianhao Fu, Jian Cao, Xinyu Zhao, Xinxin Xu, Xixin Cao, Xing Zhang

Abstract: Highlights•We find SSL in point cloud is trapped in spatial information and entangled features.•We propose GPICTURE to incorporate high-level features as additional pretext tasks.•We use CHR and DGPL to weight the low-level and high-level reconstruction losses.•We conduct a theoretical analysis from the perspective of mutual information.