Unsupervised depth estimation from monocular videos with hybrid geometric-refined loss and contextual attention

摘要

Most existing methods based on convolutional neural networks (CNNs) are supervised, which require a large amount of ground-truth data for training. Recently, some unsupervised methods utilize stereo image pairs as input by transforming depth estimation into a view synthesis problem, but need stereo camera as an additional equipment for data acquisition. Therefore, we use more available monocular videos captured from monocular camera as our input, and propose an unsupervised learning framework to predict scene depth maps from monocular video frames. First, we design a novel unsupervised hybrid geometricrefined loss, which can explicitly explore more accurate geometric relationship between the input color image and the predicted depth map, and preserve depth boundaries and fine structures in depth maps. Then, we design a contextual attention module to capture nonlocal dependencies along the spatial and channel dimensions in a dual path, which can improve the ability of feature representation and further preserve fine depth details. In addition, we also utilize an adversarial loss to discriminate synthetic or realistic color images by training a discriminator so as to produce realistic results. Experimental results demonstrate that the proposed framework achieves comparable or even better results than those trained with monocular videos or stereo image pairs. (C) 2019 Elsevier B.V. All rights reserved.