• CV
• Data
• Publications
• Software
• Teaching
• Talks

Student Project 2008-2009

Efficient and effective representations for image categorization

Student: Gaspard Jankowiak
Supervisor : Jakob Verbeek

Context:

This project is focused on image categorization, that is to tell whether an image contains one or more object categories, such as bicycles, cars, people, dogs, etc. Image categorization finds applications in image and video search and indexing. The challenge is to create categorization systems automatically from a collection of training images, for which it is known which categories they contain. Current state-of-the-art image categorization systems (as developed by Lear) represent an image as collection of local features, that describe the content of small image patches. The local features are clustered, so that each patch can be represented by the cluster to which it belongs. The clustering is often referred to as a visual vocabulary, as patches in an image represented in terms of the groups are in a sense similar to words in a text. Images are then represented by a histogram of visual word occurrences. The histograms are then used to classify the images. Two drawbacks of this approach are that (i) the process to find the group to which a patch belongs is costly (linear cost in the number of groups), and (ii) the groups are formed in a generic way that does not take into account the subsequent categorization task, and may thus be sub-optimal.

Goal:

The goal of this project is to develop new fast hierarchical clustering methods (logarithmic cost in the number of clusters), which are optimized for the final categorization task. Previous work in LEAR has shown that decision trees provide a fast and often more effective alternative. In this project we want to investigate the performance of new criteria to optimize the decision trees.
Thorough evaluation is required, comparing to alternative clustering methods, including recent methods based on Fischer Kernels. Furthermore, similarity metrics (or "kernels") between images based on the clustering results need to be evaluated, such similarity metrics are used in many current state-of-the-art image categorization methods.