Unsupervised single-shot depth estimation using perceptual reconstruction

Abstract Real-time estimation of actual object depth is an essential module for various autonomous system tasks such as 3D reconstruction, scene understanding and condition assessment. During the last decade machine learning, extensive deployment deep learning methods to computer vision has yielded approaches that succeed in achieving realistic synthesis out a simple RGB modality. Most these models are based on paired RGB-depth data and/or availability video sequences stereo images. However, lack pairs, sequences, or images makes challenging task needs be explored more detail. This study builds recent advances field generative neural networks order establish fully unsupervised single-shot estimation. Two generators RGB-to-depth depth-to-RGB transfer implemented simultaneously optimized using Wasserstein-1 distance, novel perceptual reconstruction term, hand-crafted image filters. We comprehensively evaluate custom-generated industrial surface set well Texas Face Recognition Database, CelebAMask-HQ database human portraits SURREAL dataset records body depth. For each evaluation dataset, proposed method shows significant increase accuracy compared state-of-the-art single-image methods.