Core Concepts
A normalizing flow model is introduced that operates on the product space of SO(3) manifolds, enabling accurate probabilistic modeling of human pose by respecting the rotational nature of human joints.
Abstract
The content presents a normalizing flow model, called HuProSO3, that is designed to learn normalized densities on the product space of SO(3) manifolds. This is crucial for accurately modeling the distribution of human poses, which are significantly influenced by the rotational nature of human joints.
The key highlights are:
The normalizing flow layers are designed to operate directly on the SO(3) manifold, using a Möbius coupling layer and a quaternion affine transformation. This ensures that the learned density is normalized on the space of rotations with three degrees of freedom.
The joint density on the product space of multiple SO(3) manifolds is learned using a nonlinear autoregressive conditioning approach, which can capture the statistical dependencies between human joints.
The effectiveness of HuProSO3 is demonstrated through various applications involving probabilistic human pose modeling, such as learning an unconditional pose prior, solving inverse kinematics with partial observations, and 2D to 3D pose uplifting. HuProSO3 outperforms several state-of-the-art methods in these tasks.
The proposed approach not only addresses the technical challenge of learning densities on SO(3) manifolds, but it also has broader implications for domains where the probabilistic regression of correlated 3D rotations is important.
Stats
The content does not provide any specific numerical data or statistics. The evaluation is based on qualitative and quantitative comparisons with other methods on various human pose modeling tasks.