Pipeline for self-supervised pretraining using 3D point clouds. (IMAGE)
Caption
The workflow for AI-supervised pretraining using 3D point clouds starts with unlabeled point clouds (a), followed by self-supervised learning using deep neural networks. The learned point cloud representations are then used in supervised downstream tasks (b) to initialize the network. Finally, the trained networks are fine-tuned using a small amount of labeled task-specific point cloud data to achieve high performance.
Credit
Computational Visual Media, Tsinghua University Press
Usage Restrictions
Image may be redistributed, with proper attribution, as part of news coverage of this paper only.
License
Original content