Indoor Scene Category Merging Using a Modified Davies-Bouldin Index
Open Access
Author:
Coelho, Anthony
Area of Honors:
Computer Science
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
Alan Richard Wagner, Thesis Supervisor Danfeng Zhang, Thesis Honors Advisor
Keywords:
Machine Learning Silhouette Index Davies-Bouldin Index Centroid Based Concept Learning Category Merging
Abstract:
This thesis presents a method of merging categories in a Centroid Based Image Classifier to increase accuracy. By modifying the Davies-Bouldin Index, this method identifies conceptually similar categories by analyzing their overlap in the feature-space of the model. The method is tested on the SUN RGB-D dataset which is comprised of depth and image data for different categories of indoor scenes. Indoor scene classification is an area of Machine Learning that is still developing due to its difficult nature and the method provided in this thesis aims to solve some of the accuracy difficulties of such a task. This thesis also introduces an approach to category merging that uses human intuition about a selection of categories to determine which to merge.