Simulation method of projected texture stereo for evaluating patterns

Abstract

Depth reconstruction such as stereo vision system or Kinect is widely used in robot applications. Recently, various researches which utilized projecting patterns are attempted to solve weakness of stereo vision system when the scene does not have enough texture. However, the research problem of pattern projecting method is difficult to evaluate performance of the result. In this paper, we propose quantitative performance measurement in the context of random dot pattern projection using a set of Middlebury images. Virtual projection pattern which is making with values of Ground Truth is overlaid on the original image. We perform stereo matching using these overlay images, and are able to obtain a quantitative performance through the Middlebury's test. Finally, a few result of simple stereo matching under various pattern and input situation shows that our simulation method has been a well-established.