3D Representation Learning

3D representation learning studies how to encode point clouds, meshes, voxel grids, implicit fields, Gaussian splats, and multi-view images into features that preserve shape, spatial relations, and meaning. Methods include masked and contrastive self-supervised learning, rotation-equivariant networks, neural fields, generative latent models, and alignment with image, text, or action features. Systems are tested on tasks such as ShapeNet object recognition and reconstruction, ScanNet scene understanding, and robotic manipulation benchmarks. A central challenge is balancing compact, efficient features with the geometric detail, topology, and robustness needed for sparse views, occlusion, and transfer across datasets or tasks.