Predicting depth from an image is an essential problem in the area of computer vision and deep learning shows a great potential in this area. However most deep Convolutional Neural Networks are need to train them using vast amount of manually labelled data, which is difficult or even scarcely possible in some special environment. In this paper, we proposed an unsupervised method based on left-right consistence with multi-loss fusion, which can perform single image depth estimation, despite the absence of ground truth data. We treat the issue as an image reconstruction problem by training our network with a combine of SSIM and Huber loss. To achieve estimation the depth from coarse to fine, we estimate a coarse map in the former layer and using bilinear sample to transmit the map to the latter layer to obtain a fine depth map. Our method achieves more accurate result on KITTI driving dataset.
Paper
Full text
From Coarse to Fine: A Monocular Depth Estimation Model Based on Left-Right Consistency
Semantic Scholar · Computer Science · 2019
Abstract
Predicting depth from an image is an essential problem in the area of computer vision and deep learning shows a great potential in this area. However most deep Convolutional Neural Networks are need to train them using vast amount of manually labelled data, which is difficult or even scarcely possible in some special environment. In this paper, we proposed an unsupervised method based on left-right consistence with multi-loss fusion, which can perform single image depth estimation, despite the absence of ground truth data. We treat the issue as an image reconstruction problem by training our network with a combine of SSIM and Huber loss. To achieve estimation the depth from coarse to fine, we estimate a coarse map in the former layer and using bilinear sample to transmit the map to the latter layer to obtain a fine depth map. Our method achieves more accurate result on KITTI driving dataset.