Classification of sports videos with combination of deep learning models and transfer learning

Abstract

Sports classification has considerable importance for digital content archiving in broadcasting companies. It is also a subdivision of human action recognition, which further contributes to understand the context of video scenes. In this work, deep neural networks are used, combining convolutional and recurrent networks to classify 15 individual sports classes. The sports dataset is hand-crafted to focus on sports action-based classification. CNN extracted features are combined with temporal information from RNN to formulate the general model to solve the problem. Later, transfer learning is applied with the VGG-16 model which was able to achieve 94% and 92 % test accuracy for 10 and 15 sports classes respectively.