Dominant Convolutional Feature Channels for Image Subcategory Clustering

Abstract

This thesis work aims to study what convolutional neural network actually learn and how can we make use of the convolutional neural network features. It carried out a framework to perform image subcategory clustering in the manner of poses or viewpoints. Our work is based on deep convolutional neural network feature maps and using Fuzzy c-means and K-means for clustering. To evaluate the result, we integrated our work with DPM detector and tested on PASCAL VOC 2007 dataset. The result shows our approach did improve the performance significantly in some of the categories, such as bottle, cat, table, sheep and TV monitor comparing with the original DPM detector.