Depth map repairing using tensor voting

Due to decrease in the cost of scanning devices, range maps (depth maps) are easy to capture. Depth maps can even be reconstructed from multiple images with multi-view stereo principle. Many a time, these depth maps suffer from missing regions due to occlusions, reflectivity, lack of point correspondences, sensor imperfections etc. To use these depth maps for subsequent applications, they need to be repaired by filling-in the missing values. Noise in real depth maps makes the problemnon-trivial. In this paper, we propose a depth-map repairing approach using Tensor Voting (TV) which automatically locates defective regions and estimate missing values. Our approach is well suited for real depth maps and can be a potential module for real time applications. We evaluated our method on real depth maps captured with Kinect range scanner as well as on those reconstructed from multiple images. Results show that our approach is quite effective in filling-in missing regions.