Activity Recognition by Delving deeper using CNN and RNN

Abstract

Video content has a protagonist role in this age of data revolution. These days, computer vision research community is fascinated towards application of convolutional neural networks for various image and video analysis tasks. Recurrent Neural Networks are also used in various computer vision tasks. Introduction of residual connections in traditional CNN model to design very deep architectures known as Residual Networks are very efficient for computer vision tasks. To exploit capabilities of both CNN and RNN the proposed model is based on CRNN which is trained from scratch as well as using ResNet 152 which is pre trained on ImageNet dataset. The architecture is trained and validated on popular UCF-101 dataset on the basis of accuracy and average loss. From results, it can be observed that proposed approach provides better results than state of art methods.