Boosting for Generic 2D/3D Object Recognition
Generic object recognition is an important function of the human visual system. For an artificial vision system to be able to emulate the human perception abilities, it should also be able to perform generic object recognition. In this thesis, we address the generic object recognition problem and present different approaches and models which tackle different aspects of this difficult problem. First, we present a model for generic 2D object recognition from complex 2D images. The model exploits only appearance-based information, in the form of a combination of texture and color cues, for binary classification of 2D object classes. Learning is accomplished in a weakly supervised manner using Boosting. However, we live in a 3D world and the ability to recognize 3D objects is very important for any vision system. Therefore, we present a model for generic recognition of 3D objects from range images. Our model makes use of a combination of simple local shape descriptors extracted from range images for recognizing 3D object categories, as shape is an important information provided by range images. Moreover, we present a novel dataset for generic object recognition that provides 2D and range images about different object classes using a Time-of-Flight (ToF) camera. As the surrounding world contains thousands of different object categories, recognizing many different object classes is important as well. Therefore, we extend our generic 3D object recognition model to deal with the multi-class learning and recognition task. Moreover, we extend the multi-class recognition model by introducing a novel model which uses a combination of appearance-based information extracted from 2D images and range-based (shape) information extracted from range images for multi-class generic 3D object recognition and promising results are obtained.