Abstract

We present MaskPOINT, a novel scheme of masked-wise pretrained models for point cloud self-supervised learning, addressing the challenges posed by 3D understanding, including the ambiguous underlying geometry and irregular topology information. In fact, surface-based geodesic topology pro- vides strong cues for 3D semantic analysis and geometric modeling, but such connectivity information is lost in point clouds and hardly considered. In view of this, we formulate this task into an point sampling problem and man- ifold understanding problem, and develop two techniques, a differentiable topology sampling module and manifold contrastive learning, to enable it. Specifically, the Topology Sampling samples a point cloud into point cloud geodesic field given a query point. Then, the Manifold Understanding enables the point cloud dived into topology strands based on manifold contrastive learning. Then, we randomly mask out some strands of input topology and feed them into the vanilla Transformers. The pre-training objective is to recover the underlying geometry and irregular topology information the masked. We evaluate our pretrained models across several downstream tasks, including point clouds classification, 3D segmentation, few-shot, registra- tion and demonstrate competitive results while representations learned by Mask Point owning strong interpretability.