Training Wasserstein GANs for Estimating Depth Maps

Abstract

Depth maps depict pixel-wise depth association with a 2D digital image. Point clouds generation and 3D surface reconstruction can be conducted by processing a depth map. Estimating a corresponding depth map from a given input image is an important and difficult task in the computer vision field. Fortunately, with the advent of artificial intelligence, and especially deep learning based techniques new approaches for difficult tasks have been developed. One of the attractive structures is named as Generative Adversarial Network (GAN). However, training a GAN has been reported to be problematic in terms of optimization leading to some convergence issues. Vanishing or exploding gradients and mode collapses are some examples of these issues. Lately, several alternative optimization functions and distance measures have been investigated in order to handle these difficulties. Among these approaches, Wasserstein-1 distance and Wasserstein GAN (WGAN) offers a promising alternative. In this study, Wasserstein functions and its variants are investigated for the depth map estimation task from a given 2D face image. Different network structures are trained and compared in order to assess the effectiveness and stability. Quantitative analysis is conducted by calculating two separate error metrics between the network outputs and ground-truth values.