Discriminative parts in computer vision : discovery and application

Abstract

Discriminative part-based approaches have become increasingly popular in the past few years. The reason of their popularity can be attributed to the fact that discriminative parts have the ability to accumulate low level features to form high level descriptors for objects and images. Unfortunately, state-of-the-art algorithms heavily rely on SVM related techniques, which consume a lot of computation resources in training. To overcome this shortage and apply discriminative part-based techniques to more complicated computer vision problems with larger datasets, a fast, simple and powerful algorithm named Fast Exemplar Clustering (FEC) is proposed in this dissertation. It can train part classifiers automatically in an extremely efficient manner with only class labels provided.

To show the great power of FEC, experiments were carried out on two computer vision topics: scene classification and scene text recognition. On scene classification, a new dataset named Outdoor Sight 20 was created and used in combination with MIT Indoor 67 dataset to test FEC’s ability to classify indoor and outdoor scenes. On scene recognition, a concrete example of integrating FEC was presented. Comparisons were made to show that the parts discovered by FEC were more meaningful than the existing linear SVM based feature pooling method, which led to a better recognition result.