Object categorization using collections of parts and second order pooling features

Abstract

This thesis presents an investigation of the Collection of Parts Model for object categorization. Multiclass categorization is performed using the Collections of Parts model. Results using Support Vector Machines, L1 Logistic Regression and Boosted Decision Trees are presented and discussed. Methods to analyze confusion in these results are developed and results are presented. The Collections of Parts model is augmented with features from features generated by Second Order Pooling resulting in a significant improvement in performance.