Browsing by Author "Newman, Gregory"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Video classification using deep learning
    (Stellenbosch : Stellenbosch University., 2020-03) Newman, Gregory; Brink, Willie; Herbst, B. M.; Stellenbosch University. Faculty of Science. Department of Mathematical Sciences (Applied Mathematics).
    ENGLISH ABSTRACT: To help analyse, classify, and monitor video data we need scalable algorithms that can handle video sequences of various lengths. Existing approaches tend to be both computationally expensive and restricted to classifying sequences of a fixed length, making them ill-suited for real-world use. For video classification we explore using convolutional neural networks to learn the spatial features relevant to each frame of a video, and several transfer learning approaches to leverage the InceptionV3 architecture with weights pretrained on ImageNet. With Grad-CAM we show that CNN models alone primarily rely on detecting class specific objects within images, and perform poorly on classes that have similar spatial features to other classes. To learn the temporal features of a video and to accommodate variable length sequences, we train LSTM and GRU networks. We show that without downsampling the frames the parameter space of the networks explodes, quickly becoming computationally infeasible to train over, but that downsampling techniques cause too much information loss. We also find comparable performance between the two types of recurrent networks, despite the GRU network having fewer parameters. We go on to propose an architecture that uses InceptionV3, with pretrained weights, to learn representations of the frames to be used when training a GRU network. After experimenting with different transfer learning approaches we show that we can achieve a top-5 classification accuracy of 91.8% on the UCF- 101 test set, which is 6.2% less than the state-of-the-art while having half as many parameters and an architecture that can accommodate variable length inputs.