Login Form
MSc Project PDF  | Print |

Project Details

The title of my MSc thesis was "Visual Object Class Recognition". I worked under the supervision of Dr Chris Williams. The objective of this project was to recognize the object classes of the training data in novel images. An example could be the following image (mouse over for regions of interest): 

An example image

 During the first semester of my MSc I did a literature review, the results of which are available in the end of this page. The second semester I submitted my research proposal, after discussing the review with my supervisor and investigation the results of the PASCAL contest.

The actual thesis, explores the possible extensions of the idea of the winning entry of contest. The original idea is an approach that gives state-of-art results, and belongs to Jianguo Zhang. A brief description could be the following:

The regions of interest (based on a detector, for instance Harris Affine Region Detector) are located in the training images. [mouse over on the example image] A local region descriptor is estimated (SIFT descriptors was my choice for a number of reasons) and then all the training data descriptors are collected. They are used to create a visual vocabulary, on basis of which the image-content is expressed.

Vocabulary Construction
A schematic representation of the vocabulary construction

Each word of the vocabulary corresponds to a cluster bin, and the descriptors extracted from an image are assigned to their closest bin. The number of descriptors assigned to each bin is used to create a histogram over the different descriptor clusters which serves us as the image descriptor. Implementation of the method is presented in detail in the thesis.

These image descriptors can be used along with any machine learning method to make inference, and the most interesting classification methods of the data mining field are being compared in this thesis. Furthermore, different methods are tested and compared with the results of the other contestants, as well as with the results of Zhang's specific approach.

Deliverables

  1. Literature Review
  2. MSc Thesis
  3. Presentation of the Thesis (I.A.S. group, Amsterdam, July 2005)

Further Work

The work of this thesis was further developed under my supervision in the MSc thesis of Derk Crezee. More details on this project can be found here, while a pre-print of the relative published paper and the presentation poster for the corresponding conference can be found on this page for coherence reasons.

Related Publications 

  1. A. K. Noulas, B. J. A. Kröse, Unsupervised Visual Object Class Recognition, Advanced School of Computing & Imaging Conference '06, Lommel, Belgium. [pre-print] [bib] [poster]
  2. D. Crezee, A. K. Noulas, Z. Zivkovic, B. J. A. Kröse, Visual Object Tracking and Recognition in a Visual Vocabulary Framework, Advanced School of Computing & Imaging Conference '07, Heijen, The Netherlands. [pre-print] [bib] [poster]

 
 
Joomla Templates by Joomlashack