Self-supervised contrastive learning has demonstrated remarkable effectiveness in learning visual representations without labeled data, yet its application to 3D local feature learning from point clouds remains underexplored. Existing methods predominantly focus on complete object shapes, neglecting the critical challenge of recognizing partial observations commonly encountered in real-world 3D perception. We propose a momentum contrastive learning framework specifically designed to learn discriminative local features from randomly sampled point cloud regions. By adapting the MoCo architecture with PointNet++ as the feature backbone, our method treats local parts of point cloud as fundamental contrastive learning units, combined with carefully designed augmentation strategies including random dropout and translation. Experiments on ShapeNet demonstrate that our approach effectively learns transferable local features and the empirical observation that approximately 30% object local part represents a practical threshold for effective learning when simulating real-world occlusion scenarios, and achieves comparable downstream classification accuracy while reducing training time by 16%.
Sha et al. (Tue,) studied this question.