CVPR 2023 – Delving into Discrete Normalizing Flows on SO(3) Manifold for Probabilistic Rotation Modeling

In this episode we discuss Delving into Discrete Normalizing Flows on SO(3) Manifold for Probabilistic Rotation Modeling
by Yulin Liu, Haoran Liu, Yingda Yin, Yang Wang, Baoquan Chen, He Wang. The paper proposes a new normalizing flow method for the SO(3) manifold, which is an important quantity in computer vision, graphics, and robotics but has unique non-Euclidean properties that make it difficult to adapt existing normalizing flows. The proposed method combines a Mobius transformation-based coupling layer and a quaternion affine transformation to effectively express arbitrary distributions on SO(3) and allows for conditional building of the target distribution given input observations. Extensive experiments show that the proposed rotation normalizing flows outperform baselines on both unconditional and conditional tasks.