Single-image depth estimation by refined segmentation and consistency reconstruction

Abstract

Recent years have witnessed tremendous success of single-image depth estimation. However, most of the existing approaches merely use scene descriptions of a whole image to retrieve its candidates, which may end up with undesirable depth supports for local regions. In this paper, we propose a segmentation method for single-image depth estimation based on data-driven framework. First, a per-pixel boundary spreading method is presented to improve the image segmentation and provide local regions for image retrieval. Second, a local-region image retrieval is conducted to provide a powerful support for the depth estimation of each segmented part. Third, a scene similarity matrix is constructed and combined with the initial depth prior to establish the correlations across different regions for a consistent depth optimization. Experiments show that applying our method to classic data-driven methods can improve the performance of depth estimation. Besides, our results also manifest clearer depth boundaries in some local regions than the state-of-the-art methods based on deep learning framework.