الفهرس | Only 14 pages are availabe for public view |
Abstract This thesis is an example of how high performance computing tech- niques can be used to speedup the processing of computationally inten- sive problems. The problem of machine ability to catgeorize dierent objects is our selected problem. Object categorization is the task of classing objects into generic classes. Although such task is eort-less for humans, it is a complex and computationally intensive for com- puters. Object categorization is an important task for its dierent applications such as image annotation, image retrieval, video anno- tation,surveillance, driver assistance, autonomous robots, interactive games. It is also used as apreparation step for object recognition. The performance metric for the categorization task consists of two key measurements, success categorization rate and system run-time. Dierent trials to solve the categorization problem are currently in place. However, few trials consider both of them. In fact, the heavy processing tasks needed for accurate categorization system lead to in- crease system run-time. Recently, the high performance techniques are used to solve the problem of increasing system run-time In this thesis, we provide a study of 3D object categorization al- gorithms based on complex feature and a study of dierent high per- formance computing techniques that could be used to enhance such algorithms performance. We introduce a system that categorizes 3D objects based on their depth information. It matches the real time constraint with high success categorization rate compared to other existing systems. In the proposed categorization system, spin-images are selected as features also Support Vector Machine is selected as a classier. Spin- image is a complex feature to be extracted. We investigated the task dependancy in such feature extraction. Moreover, we modied the original spin-image algorithm to eliminate the unnecessary blocking tasks based on our task dependancy analysis. Dierent strategies enhancing spin-point selection in order to en- hance sucess categorization rate are evaluated. Also, we evaluated the benift of using bag of features technique for the success categoriza- tion rate. During our initial experiments, the success categorization rate was approached 65%. After futher improvements, this rate has been signicantly increased such that we have achieved almost 81% in 0.85033 second for each single object. Eventually, an evaluation to our implemented categorization sys- tem is carried out in comparison to two dierent types of publicly available 3D objects datasets. The rst one was the Princeton shape benchmark. The second one was the RGB-D dataset. Results have proved that our categorization system provides much more accurate and faster categorization. |