A Semi-Supervised GAN Architecture for Video Classification

Abstract

In recent years, several supervised deep-learning architectures have achieved state-of-the art accuracies in video-classification. However, they demand a considerable amount of annotated data which can be both cost and resource intensive. This study proposes a Semi-Supervised GAN architecture to efficiently perform classification on video datasets with a small percentage of labelled data. While the Generative Adversarial Network (GAN) architecture is known for its generative ability, we harness the discriminative property of this network instead for the classification of videos. The proposed model leverages the features extracted from the unlabelled data to classify the labelled videos. Results show that the proposed approach achieves 46% accuracy with just 5% labelled videos, reaching up to 62% when 50% of the videos are labelled. These results are a significant improvement over a standard supervised approach and show a promising aspect in the field of Semi-Supervised Learning domain.