Research on Monocular Depth Estimation Algorithm Based on Structured Loss

Abstract

This paper proposes a monocular depth estimation algorithm based on multi-scale structure similarity and gradient matching for improving the accuracy of monocular image depth estimation and solving the problems of inaccurate prediction of geometric shapes and blurred edges in the image. In this algorithm, a joint structured loss is formed by using multi-scale structure similarity degree loss and scale-invariant gradient matching loss. The relative depth points are sorted to achieve monocular depth estimation, which realizes accurate prediction of geometric shapes in the image, reduces edge blur, and improves depth prediction accuracy. Numerical experiments and subjective evaluations are performed on four different types of data sets: Ibims, NYUDv2, DIODE, and Sintel. The results show that the algorithm significantly reduces the depth prediction error, effectively improves the accuracy of the prediction, and has a certain generalization performance.