Using Convolutional Neural Networks with Transfer Learning for Action Recognition in Videos

Abstract

Video classification is a broad field within the machine learning landscape. There are many videos on the internet and being able to classify them allows people to keep track of certain trends. Due to the volume of videos this task is a challenge. In this research we design models on a relatively small dataset to explore the possibilities of separating action recognition from object recognition in hopes of improving the accuracy of previous models. We also explore data augmentation techniques so as to try and expand a limited dataset. The models that we designed are able to identify features from motion data, which allows it to perform better at action recognition.