Effective Color Spaces for Quaternion-valued Neural Network in Depth Estimation

Abstract

In the development of deep learning technology, we frequently focus on how to create the best neutral architecture to enhance models and obtain higher accuracy while overlooking a way to speed up training because any parameters are affected by color space. Finding the ideal color space for Quaternion-valued neural network in-depth estimation as survey methods is the focus of this paper. We use a small dataset from the Middlebury dataset [1] to survey training progress in a quaternion-valued neural network that was mentioned in one of our other papers. As a result, we find that HED color-space makes the best training progress in the survey results.