Unsupervised Stereo Depth Estimation Refined by Perceptual Loss

Abstract

Depth of the object has long been a critical information in mobile robot filed and computer vision. In recent years, binocular depth estimation based on supervised learning with deep convolutional neural network has seen huge success when compared with traditional or unsupervised methods. Despite all this, unsupervised depth estimation methods still need further study because they conquer the vast quantities collection of corresponding ground truth depth data for training. To resolve this, methods based on semi-supervised learning are proposed, where stereo images are reconstructed according to predicted disparities. Compared with supervised learning, the maximum restriction is the ill-posed problem of image color similarity between the reconstructed image and the input color image. To improve this problem, in this paper we combine the more robust perceptual loss with image color loss to encourage the similarity between the images feature representations extracted from another convolutional neural network. Benefited of the both losses, we improve the stereo depth estimation accuracy proposed by Godard et al. on KITTI benchmark.