Self-Supervised Representation Learning for CAD

Few-shot Task Learning Task Decoder Pretrained Encoder Small Labeled Dataset Encoder Decoder u,v d x,y,z B-Rep Model Large Unlabeled Dataset Face Embeddings Rasterization Figure 1. Overview of our technique. We train a geometric self-supervision task of a large, unlabeled dataset of CAD Boundary Representations (B-Reps) to learn geometrically relevant representations for each B-Rep face. These pre-trained representations are then used to train few-shot segmentation and classification learning tasks on labeled B-Rep datasets.