Exploiting data expansion and generation for visual recognition based on deep models

Abstract

This research is motivated by the question: When data is scare, how can we use data expansion methods or generative models to learn better representations of deep models for visual recognition? To solve the question, we propose a Generative Translation Classification Networks (GTCN) that employ joint learning to train a classifier and a generative stochastic translation network end-to-end. The methods are researched on face liveness detection, cat/dog classification, and artist classification as experimental applications.