Point-to-Block Matching in Depth Estimation

Abstract

In this paper, the novel correspondence search method called point-to-block matching was proposed. Recently, in many proposed multimedia systems, depth estimation is performed on compressed input views. To address this problem and increase the quality of such depth maps, in the proposal, a point in a view is not compared simply with a point in another view, but to the most similar point in a small block surrounding it. The introduction of this method in the depth estimation process is beneficial to the quality of depth maps, as it decreases the influence of small shifts in images, caused e.g., by encoding-related errors introduced to input views. The method was implemented in one of the state-of-the-art depth estimation methods and tested in series of experiments. Based on the comparison of synthesized virtual views with the input views, the proposal increases the quality of estimated depth maps in most of the tested configurations.