Point-DMAE: Point Cloud Self-supervised Learning via Density-directed Masked Autoencoders

Masked autoencoders have been extensively utilized in 3D point cloud self-supervised learning, where the fundamental approach involves masking a portion of the point cloud and subsequently reconstructing it. This process is hypothesized to enhance model learning by leveraging the inherent structure of the point cloud data. However, the information density within point clouds is inherently uneven, contrasting with the more uniform distributions found in language and 2D image data. This uneven distribution suggests that the application of random masking strategies, commonly adopted from NLP and 2D vision, may not be optimal for point cloud data, potentially leading to suboptimal learning outcomes. Based on this observation, we propose a simple yet effective Density-directed Masked Autoencoders for Point Cloud Self-supervised Learning (Point-DMAE), which learns latent semantic point cloud features using a density-directed masking strategy. Specifically, our method employs a dual-branch Transformer architecture to extract both high-level and fine-grained point features through global and local block density-directed masking, respectively. Point-DMAE demonstrates high pre-training efficiency and significantly outperforms our baseline (Point-MAE) on 3D object classification tasks within the ScanObjectNN dataset by 4.13% on OBJ-BG, 5.17% on OBJ-ONLY, and 4.17% on PB-T50-RS. Codes are available at https://github.com/jinxianglong10/Point-DMAE.