Masked Discrimination for Self-supervised Learning on Point Clouds

Masked autoencoding has achieved great success for self-supervised learning in the image and language domains. However, mask based pretraining yet to show benefits point cloud understanding, likely due standard backbones like PointNet being unable properly handle training versus testing distribution mismatch introduced by masking during training. In this paper, we bridge gap proposing a discriminative Transformer framework, MaskPoint, clouds. Our key idea is represent as discrete occupancy values (1 if part of cloud; 0 not), perform simple binary classification between masked object points sampled noise proxy task. way, our approach robust sampling variance clouds, facilitates rich representations. We evaluate pretrained models across several downstream tasks, including 3D shape classification, segmentation, real-word detection, demonstrate state-of-the-art results while achieving significant speedup (e.g., 4.1 $$\times $$ on ScanNet) compared prior baseline. Code available at https://github.com/haotian-liu/MaskPoint .