Residual Feature Pyramid Architecture for Monocular Depth Estimation

摘要

This paper investigates Visualization of Image Depth Information architecture of fully convolutional residual networks, based on it, residual feature pyramid network architecture for monocular image depth estimation is proposed. Firstly, the input monocular RGB image is preprocessed. Secondly, the feature pyramid structures are introduced into the fully convolutional residual networks, which realizes multi-scale feature extraction and reused, at the same time, both the number of parameters in networks and the computational complexity are greatly reduced. Finally, Experiments are done on the commonly used NYU official dataset. Experimental results show that the proposed method has advantages over many recent advanced methods. And the object outlines in our inferred depth maps are clearer and exquisite which look qualitatively better. In addition, this paper uses a simpler networks structure to realize a lower system error.