Depth Extraction from Monocular Video Using Bidirectional Energy Minimization and Initial Depth Segmentation

TL;DR: The bidirectional energy minimization in this scheme considers both the previous frame and next frame, which promises a much more robust depth map and reduces the problems associated with occlusion to a certain extent.

Abstract: In this paper, we propose to extract depth information from a monocular video sequence. When estimating the depth of the current frame, the bidirectional energy minimization in our scheme considers both the previous frame and next frame, which promises a much more robust depth map and reduces the problems associated with occlusion to a certain extent. After getting an initial depth map from bidirectional energy minimization, we further refine the depth map using segmentation by assuming similar depth values in one segmented region. Different from other segmentation algorithms, we use initial depth information together with the original color image to get more reliable segmented regions. Finally, detecting the sky region using a dark channel prior is employed to correct some possibly wrong depth values for outdoor video. The experimental results are much more accurate compared with the state-of-the-art algorithms.