Master Thesis

6D Object Instance Recognition from RGB images: A Learning-Based Perspective

Work default illustration


  • Started: 01/03/2018


6D model-based rigid object pose retrieval from vision is an important problem in computer vision. Early solutions address it as an instance recognition problem, by means of template matching approaches. In contrast, in this work we introduce a learning-based approach. To this end, the projection space is defined by using an own-developed descriptor that it is then optimally segmented. After that, we make use of a probabilistic model to match every given instance to its belonging segment. Thus, obtaining a valid subspace for its dominion. The method only requires RGB information as input, as well as a full CAD model of the observed objects. It intends to be adequate for texture-less objects. Robustness to occlusion and clutter are desirable attributes for the method, also in the scope of the project.