The Continuity of Rotation Representations in Neural Networks [pdf] (openaccess.thecvf.com)

A recent study from researchers at the University of Southern California and Adobe Research has introduced a new framework for continuous representations of rotations in neural networks, challenging traditional methods that utilize quaternions and Euler angles. The authors found that while these common representations are discontinuous and lead to significant regression errors during training, transitioning to continuous shapes in 5D and 6D allows for better performance in tasks such as 3D rotation estimation and inverse kinematics. By establishing a definition of continuity based on topological concepts, the research highlights that neural networks can learn with lower approximation errors when working with continuous representations. This work is significant for the AI/ML community as it addresses the inherent challenges presented by discontinuous representations in neural network training, particularly for applications in graphics and vision, where accuracy in rotation is crucial. The theoretical insights and empirical evidence provided suggest that adopting continuous representations can substantially enhance the reliability and efficiency of neural networks, potentially paving the way for more robust AI solutions in dynamic environments. This advancement is poised to improve techniques in motion capturing, pose estimation, and other related fields, thereby expanding the practical applications of deep learning in real-world scenarios.