Cluster-MAE: Enhancement of Mask Autoencoder for Point Cloud Big Data

Point clouds have become a crucial data representation for 3D vision applications such as autonomous driving and robotics, fueled by advancements in deep learning. Advances in 3D acquisition technologies, such as LiDAR, have simplified the collection of massive point cloud datasets. Effectively utilizing the vast quantities of unlabeled point cloud data necessitates efficient and robust learning methods. Inspired by the success of masked autoencoders (MAE) in natural language processing and computer vision, recent research has explored their application to point cloud processing. However, existing MAE-based point cloud pre-training methods suffer from inefficient point cloud partitioning, resulting in both data loss and redundant computations. This inefficiency hinders their applicability to large-scale point cloud data. This paper introduces Cluster-MAE, a novel and efficient self-supervised pre-training paradigm for point clouds. Cluster-MAE incorporates a novel clustering strategy for generating non-overlapping point clusters, minimizing computational overhead, and ensuring complete data utilization. Furthermore, a dedicated cluster interaction module enhances the model's ability to learn rich representations. Extensive experimental results demonstrate the effectiveness of Cluster-MAE, achieving competitive accuracy with a significant reduction in computational resources $(62 \%-83 \%)$ and memory requirements (59%) compared to previous methods. This makes it ideally suited for point cloud big data.