Well Convergent and Computationally Efficient Quaternion Loss

摘要

Rotation estimation, i.e. the ability to predict angles describing 3D positioned object, is an omnipresent problem in computer vision, computer graphics and 3D object detection task in automotive industry. Deep learning algorithms usually parameterise rotation using only the yaw angle. This paper presents detailed comparison of several 3D rotation distance functions using quaternions. We propose a computationally efficient quaternion loss function for neural network training. We conclude that function respects the topology of SO(3) and is bi-invariant. We also show the geometrical representation of presented functions. Lastly, we evaluate the effectiveness of the proposed loss function and compare its performance with other methods using a public large-scale dataset.