We propose a self-supervised capsule architecture for 3D point clouds. We compute capsule decompositions of objects through permutation-equivariant attention, and self-supervise the process by training with pairs of randomly rotated objects. Our key idea is to aggregate the attention masks into semantic keypoints, and use these to supervise a decomposition that satisfies the capsule invariance/equivariance properties. This not only enables the training of a semantically consistent decomposition, but also allows us to learn a canonicalization operation that enables object-centric reasoning. To train our neural network we require neither classification labels nor manually-aligned training datasets. Yet, by learning an object-centric representation in a self-supervised manner, our method outperforms the state-of-the-art on 3D point cloud reconstruction, canonicalization, and unsupervised classification.
@article{arxiv.2012.04718,
title = {Canonical Capsules: Self-Supervised Capsules in Canonical Pose},
author = {Weiwei Sun and Andrea Tagliasacchi and Boyang Deng and Sara Sabour and Soroosh Yazdani and Geoffrey Hinton and Kwang Moo Yi},
journal= {arXiv preprint arXiv:2012.04718},
year = {2021}
}
Comments
NeurIPS 2021; The first two authors contributed equally; Project website: https://canonical-capsules.github.io