Improving Depth Maps by Nonlinear Diffusion

Jianfeng Yin and Jeremy R. Cooperstock
McGill University
Centre for Intelligent Machines
H3A 2A7 Montreal

e-mail: jfyin,


Dense depth maps, typically produced by stereo algorithms, are essential for various computer vision applications. For general configurations in which the cameras are not necessarily parallel or close together, it often proves difficult to obtain reasonable results for complex scenes, in particular in occluded or textureless regions. To improve the depth map in such regions, we propose a post-processing method and illustrate its benefits to applications such as 3D reconstruction or foreground segmentation of persons in a scene.