PhD Proposal: A Non-Parametric Approach to Extending Generic Binary Classifiers for Multi-Classification

Talk
Venkataraman Santhanam
Time: 
06.05.2015 10:30 to 12:00
Location: 

AVW 4424

Object recognition and verification are fundamental problems in computer vision. Several state-of-the-art computer vision and pattern recognition systems designed to solve these problems often rely on multi-classification as one of the key-components in their pipeline. One of the many ways to solve the multi-classification sub-problem is via "ensemble methods", which combine generic binary classifier scores to generate a multi-classification output. Ensemble methods are frequently used for complex multi-class problems with a large number of classes and/or high feature dimensionality, where dedicated multi-class classifiers are often computationally intractable.
In this proposal, we present a robust multi-classification pipeline that makes best use of all the relevant information encapsulated within the ensemble of binary classifier scores and provides a consistent probabilistic multi-classification output. Our approach, at a high level, involves projecting the binary classifier scores into compact orthogonal subspaces, followed by a non-linear probabilistic multi-classification step, using Kernel Density Estimation (KDE). We test our final approach on 6 multi-category computer vision datasets (MIT67, UCF50, MNIST, Birds, Butterflies, Robotics), for 2 choices of binary classifiers (Linear SVM and Composite Discriminant Factors), comparing it with state-of-the-art ensemble methods (VOTE, NEST). Experimental results show that our approach improves multi-classification accuracies over state-of-the-art for both Linear SVM and Composite Discriminate Factors.
For future work, we will focus on a more specific ORIS/OVIS (Object Recognition/Verification based on Image Set) problem, in which the test data is grouped into disjoint sets of test samples, with apriori knowledge that each test group corresponds to one object identity. One of the instances where such a scenario occurs is when test samples are extracted from videos and each test group corresponds to a collection of frames in one video. The key challenge in this problem is how to make best use of the joint information that may be obtained from the test group as a whole, a problem that is further hindered by the fact that each test group might have a potentially different cardinality. A potential solution to this problem is to represent each test group as a manifold and exploring techniques to compute distance measures between manifolds of potentially variable dimensions. Finally, we plan to work on semi-supervised/unsupervised outlier detection techniques which would make generic object recognition/verification pipelines more robust.
Examining Committee:
Committee Chair: - Dr. Larry Davis
Dept's Representative - Dr. Ramani Duraiswamy
Committee Member(s): - Dr. David Jacobs