Mutual information-driven self-supervised point cloud pre-training

Learning universal representations from unlabeled 3D point clouds is essential to improve the generalization and safety of autonomous driving. Generative self-supervised point cloud pre-training with low-level features as pretext tasks is a mainstream paradigm. However, from the perspective of mutual information, this approach is constrained by spatial information and entangled representations. In this study, we propose a generalized generative self-supervised point cloud pre-training framework called GPICTURE. High-level features were used as an additional pretext task to enhance the understanding of semantic information. Considering the varying difficulties caused by the discrimination of voxel features, we designed inter-class and intra-class discrimination-guided masking (I${}^2$Mask) to set the masking ratio adaptively. Furthermore, to ensure a hierarchical and stable reconstruction process, centered kernel alignment-guided hierarchical reconstruction and differential-gated progressive learning were employed to control multiple reconstruction tasks. Complete theoretical analyses demonstrated that high-level features can enhance the mutual information between latent features and high-level features, as well as the input point cloud. On Waymo, nuScenes, and SemanticKITTI, we achieved a 75.55% mAP for 3D object detection, 79.7% mIoU for 3D semantic segmentation, and 18.8% mIoU for occupancy prediction. Specifically, with only 50% of the fine-tuning data required, the performance of GPICURE was close to that of training from scratch with 100% of the fine-tuning data. In addition, consistent visualization with downstream tasks and a 57% reduction in weight disparity demonstrated a better fine-tuning starting point. The project page is hosted at https://gpicture-page.github.io/.