Abstract:
In this thesis, we have investigated the 3D object class recognition problem. We used an approach that solves this problem with the use of depth images obtained from 3D object models. In the approach we used, 3D object class recognition system is composed of two stages; training and testing. In both stages, rst, keypoints are detected from the images, and then 2D local image descriptors are built around these keypoints. This is continued by encoding local descriptors into a single descriptor. Just before this step, in training stage, a codebook is learned, and it is used for encoding local descriptors in both stages. Another extra step in training stage is, after the descriptors are encoded, for each class a binary classi er is trained. Then, these classi ers are used in testing stage. We have evaluated di erent keypoint detection methods, 2D local image descriptors and encoding methods. Then, we experimentally show their superiorities and weaknesses over each other. Our experiments clearly show the best performing keypoint detection method, local image description method and feature encoding method in the depth image domain, which are densely sampled SIFT descriptors and Fisher Vector encoding. Using di erent experimental setups yields similar results, thus the validity of the methods that are selected as best is proven.
