URNet: An UNet-Based Model with Residual Mechanism for Monocular Depth Estimation

Abstract

Autonomous vehicle systems rely heavily upon depth estimation, which facilitates the improvement of precision and stability in automated decision-making systems. Noteworthily, the technique of monocular depth estimation is critical for one of these feasible implementations. In the area of segmentation of medical images, UNet is a well-known encoder–decoder structure. Moreover, several studies have proven its further potential for monocular depth estimation. Similarly, based on UNet, we aim to propose a novel model of monocular depth estimation, which is constructed from the benefits of classical UNet and residual learning mechanisms and named URNet. Particularly, we employ the KITTI dataset in conjunction with the Eigen split strategy to determine the efficacy of our model. Compared with other studies, our URNet is significantly better, on the basis of higher the precision and lower error rate. Hence, it can deal properly with the depth estimation issue for autonomous driving systems.

SAU-Net: Monocular Depth Estimation Combining Multi-Scale Features and Attention Mechanisms

“…This model achieves a good balance between accuracy and computation time. Duong et al [32] constructed a depth estimation model called UR-Net, which adds attention to the decoder and replaces the transmission block in the conventional U-Net with spatial pyramid pool blocks (ASPP).…”

Monocular Depth Estimation for 3D Map Construction at Underground Parking Structures

“…He and his scholars do this by initially generating a rough prediction of deep information, and then using another neural network to refine it to produce more accurate results. Since then, a great deal of work has been done to improve the accuracy of supervised depth estimation of monocular images, including the addition of residual mechanism [16], the use of conditional random scene methods [17], the anti-Huber distance loss function method [18], the joint optimization of surface normals [19], the fusion of multiple depth maps [20], the add of an additional channel to the output layer [21] and the method of expressing it as an ordinal classification problem [22]. People even apply the method to microscopic scenes [23].…”

Comparing U-Net based architectures in monocular depth estimation

“…Devido a sua arquitetura flexível e facilmente modificável, a U-Net tem sido uma arquitetura de CNN comumente adotada como base para a construc ¸ão de novas redes na tarefa de estimativa de profundidade [29,30,31]. Saxena et al [31] combinaram a U-Net, utilizando a EfficientNet como codificador, com a reconstruc ¸ão de mapas de profundidades ruidosos para melhorar significativamente a estimativa de profundidade.…”