Self-supervised masked occupancy encoder for 3D object detection

Because of the sparse distribution and lack of organization in point clouds, 3D object detection has been a challenging task. Current detection models heavily rely on 3D labeled data, while in real-world complex scenarios, issues such as point cloud sparsity and occlusions make feature extraction difficult. In this study, we propose a pre-training approach based on self-supervised masking. We compute the distance between voxels and the LiDAR device to regulate the masking ratio of voxels, randomly masking voxels accordingly. Predicting the occupancy status of voxels enables the prediction of the masked occupancy structure for the entire 3D scene. This masking strategy enables self-supervised model training, achieving the effect of data augmentation. We conducted thorough experiments using the KITTI datasets, specifically targeting the 3D object detection task. We achieved a significant +2.11% improvement in the Hard category of Car detection. mAP across three difficulty levels also increased by +0.74%, illustrating the efficiency of our method in handling complex circumstances.