Using 2D image texture to improve 3D point clouds from stereo vision

Stereo vision systems suffer from noise and conscious steps need to be taken to process this information. One aspect of my research deals with this. In disparity images there is an inherent disadvantage in depth information as disparity is always an integer value. This creates discontinuous steps in 3D point clouds generated using these. To counter this, I am using 2D texture information to improve the disparity images. I first implement SLIC super-pixel clustering over Lab color space 2D image. Then combine similar clusters. Then I use these cluster locations on disparity image. The fit a plane on the clusters in the disparity image. What this fitting does is to provide sub-pixel data information in the disparity image. I use this to convert into 3D point cloud and get a very sharp cloud.

 

 

 

Basic tests reveal the depth information remains conserved.