Depth from motion using critical point filters with unconstraint camera motion

Depth estimation is a crucial step for 2D/3D conversion from monoscopic video. In this paper, a novel method for depth estimation from motion with camera motion is proposed. In the proposed method, image matching using critical point filters is applied to extract the pixel-level motion field for each frame. As camera motion can bring pseudo motion vectors by image matching, and thus leading to depth ambiguity. To solve this problem, we propose to estimate the camera moving model using robust RANSAC algorithm. Then, the initial depth map is estimated by using the motion vectors without camera motion. Finally, the depth values of the pixels at the edges of moving objects are refined using a post filter based on homogeneous points. Experimental results show that the proposed method achieves considerable performances on depth map in presence of camera motion.