Author:

Keywords:

3D shape analysis, recognition, retreival, segmentation, PSI_VISICS

Abstract:

This thesis focuses on three main topics in the area of 3D recognition: shape classification, retrieval, and scene segmentation. For classification, in contrast to previous methods that focused on the overall shape appearance only, we extended the popular SURF features into a third dimension and presented a novel method to learn the efficient 3D Implicit Shape Model (3D ISM), which is a class-specific star model that combines the appearance and the relative position of local patches from 3D shapes. The proposed shape retrieval approach is based on our previous classification model. Thus, 3D ISM introduces additional spatial constraints to improve the basic bag-of-visual-words ordering. This ISM-based verification step allows us to use shape expansion that gains even better results. In addition, we show how to usenbsp;retrieval pipeline to improve shape matching, classification, and text-based 3D shape search. The method was also practically used in the 3D Coform project in a search for artifacts in a museum’s repository. The previously introduced method for classification was only applied to recognize classes of shapes that are isolated, clean, and without holes. Instead, segmenting large 3D scenes is a more realistic scenario. Objects need to be firstly found and then segmented. We propose two methods. The first, combines our previous findings in classification with CRF optimization and the second is a more sophisticated framework that solves ISM and graph optimization jointly. Once the object is found and segmented from its background, we use our method (3D ISM) or introduced new Boltzmann Machines-based approach to fill-in holes and to correct the wrong parts of incomplete shape. Each task is evaluated on several benchmarks against a variety of state-of-the-art methods.