Uncertainty Guided Self-Supervised Monocular Depth Estimation Based on Monte Carlo Method

Depth estimation is an important research topic in the field of computer vision. Recently, self-supervised methods for depth estimation have received significant attention due to their independence from costly annotations. Recently, many methods for constructing uncertainty maps have been proposed to improve the performance of dense prediction tasks, such as semantic segmentation, and deep depth estimation also belongs to this category. This paper employs Monte Carlo methods to generate uncertainty maps that guide self-supervised monocular depth estimation learning. We sample multiple model outputs through the Monte Carlo method and calculate the variance and mean of the multiple output results. The mean is used as a teacher model to provide a supervisory signal, and the variance result is used as an uncertainty map of the supervisory signal. We use this uncertainty map to correct the supervision of depth information for better model performance. In this paper, we tested the challenging dataset KITTI and conducted generalization experiments on the Make3D dataset. The experimental results show that our proposed method has a significant improvement in self-supervised monocular depth estimation.