Stereo Refinement for Photo Editing

摘要

We present a method for refining depth information generated by a stereo-matching algorithm with the goal to provide depth-aware photo-effect applications. Our key idea is to use structural features of the base image to enhance the depth information. Our method pre-processes the original disparity map by revising the sky region and removing incorrect data (on the left-side of the disparity map) caused by occlusion. The base image is mean-shift segmented. A median filter is applied on the disparity map within each segment. Invalid step-edges in the disparity map are removed by a joint bilateral filter. Experiments show that our method can revise holes, inaccurate object edges, speckle noises, and invalid step-edges from the given depth information. Results illustrate the applicability for photo editing.