Scene Segmentation Assisted by Stereo Vision

Abstract

Stereo vision systems for 3D reconstruction have been deeply studied and are nowadays capable to provide a reasonably accurate estimate of the 3D geometry of a framed scene. They are commonly used to merely extract the 3D structure of the scene. However, a great variety of applications is not interested in the geometry itself, but rather in scene analysis operations, among which scene segmentation is a very important one. Classically, scene segmentation has been tackled by means of color information only, but it turns out to be a badly conditioned image processing operation which remains very challenging. This paper proposes a new framework for scene segmentation where color information is assisted by 3D geometry data, obtained by stereo vision techniques. This approach resembles in some way what happens inside our brain, where the two different views coming from the eyes are used to recognize the various object in the scene and by exploiting a pair of images instead of just one allows to greatly improve the segmentation quality and robustness. Clearly the performance of the approach is dependent on the specific stereo vision algorithm used in order to extract the geometry information. This paper investigates which stereo vision algorithms are best suited to this kind of analysis. Experimental results confirm the effectiveness of the proposed framework and allow to properly rank stereo vision systems on the basis of their performances when applied to the scene segmentation problem.