Fusing Color and Geometry Information for Understanding Cluttered Scenes

In this paper, we introduce a new image processing pipeline for scene recognition and pose estimation in robotic applications. Unknown objects are autonomously modeled resulting in geometric 3D models and color images. Theses models are then used for object recognition in cluttered scenes by merging color and geometry information. Our recognition approach generates new suitable feature vectors and uses RANSAC to obtain promising hypotheses of recognized object poses for the scene. RANSAC is widely used for scene understanding. For making RANSAC applicable, it is very important to implement this algorithm efficiently and to reject hypotheses as early as possible in the scene understanding pipeline. By using color information many hypotheses can be rejected early in the recognition pipeline. With our approach we provide an efficient implementation of a scene analyzing pipeline while fusing color and geometric information. Moreover, we are able to learn new objects by a fast autonomous scanning process and no further runs through time consuming learning algorithms are necessary. The complete pipeline from scanning to scene understanding is described. The evaluated scenes consist of several household objects. Some of them vary only in texture and not in shape.