M.Sc Thesis

M.Sc StudentBar-Aviv Ezer
SubjectFunction Based Object Categorization Using Simulation
of an Agent
DepartmentDepartment of Computer Science
Supervisor PROF. Ehud Rivlin
Full Thesis textFull thesis text - English Version


Encouraged by recent cognitive psychology studies, this research presents a novel  functional approach for the categorization of 3D objects. The work present new approach for representing functionalities by means of agent and functions, in harmony with motor cognition psychology. By embodying the receiver of the functionality within a virtual environment and simulating its interaction with a candidate object, functional properties of the object can be revealed and validated. We present a generic search algorithm, which is used during the simulation process, designed

to search the configuration space looking for evidence configurations that prove the fulfilment of the functionality. The algorithm involve two steps - searching for semi functional configurations and relaxation. The first step is targeted to accelerate the search in the complete configuration space of both global-DOFs and inner-DOFs via a unique pruning. This pruning is inspired by cognitive motives and implemented by means of collision detection. The second step is the relaxation phase which simulate the action performed by the agent upon the candidate object. The action is represented using a sequential movements of the inner-DOFs of the agent until the goal configuration is

reached. The use of an agent in the categorization process enables us to overcome the shape-diversity of objects that belong to the same category and to handle noise and self occlusions that occur in range scans. To demonstrate the robustness of the approach we have implemented the categorization algorithm for several categories and functionalities. We have tested it on a large database of both 3D-CAD graphic models and real 3D scanned data

generated from a single view point. Sensitivity analysis shows that the concept is applicable and robust enough to be used in real applications.