Two SVM-Based Methods for Image Set Classification

Abstract

Set classification problem has arisen from a number of real-life scenarios, such as video face recognition and object categorization. Different from the single observation based classification problem, in set classification, we have a priori information that all the observations in one set belong to the same type. To properly exploit this set structure, we proposed two new approaches on the basis of support vector machine, which we refer to as unity prediction and homogeneity prediction. For the unity prediction, we use the averaged predicted values for all the observations in one set as the predicted value for the set and then conduct the SVM on a set level instead of observation level. In homogeneity prediction, we utilize the set structure through imposing constraints on the predicted value to enforce that the predicted values for all the observations in one set share the same sign. These two approaches incorporated the set information in both the learning and prediction phase. Experimental evaluation of these two methods are performed on ETH-80 dataset, YouTube Celebrity dataset and Liver Cell Nucleus dataset. Comparison with the existing techniques shows that our methods constantly achieve better results.

Date of Award	16 Jul 2020
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Junhui WANG (Supervisor)