NOTICE OF 1st PhD SEMINAR PRESENTATION
CANDIDATE: Aditya Subramani Murugan
DEGREE SOUGHT: PhD
DATE: 3/16/2023
TIME: 3:00pm
PLACE: Room 2100 CEI
TITLE: Approaches to action recognition on UCF101
Abstract
Action recognition plays a pivotal role across various domains, and the accurate classification of human actions from visual data holds significant importance for real-world applications. This study conducted an in-depth comparison of three machine learning models: 3D Convolutional Neural Networks (3DCNN), Temporal Convolutional Networks (TCN), and a combination of Residual Networks (ResNet) and Long Short-Term Memory (LSTM) networks, targeting the task of action recognition on the UCF101 dataset. Leveraging a subset of the UCF101 dataset, six experiments were executed to evaluate the performance and computational efficiency of the different architectures. Each model was implemented in TensorFlow, fit for 100 epochs with early stopping, and then evaluated using a test dataset. Through accuracy-loss graphs, confusion matrices, and computations of train, validation, and test accuracies, we found that the ResNet+LSTM combination performed most effectively, achieving a 98% accuracy on the test dataset. Moreover, no significant performance difference was observed between the use of ResNet50 and ResNet152 within the combined model. Therefore, we recommend the deployment of ResNet50+LSTM for action recognition tasks due to its high accuracy and computational efficiency. Future work should focus on model stress-testing with diverse datasets, exploring transfer learning techniques, and investigating potential benefits from the fusion of different architectures. This study also points out the absence of hyperparameter optimization and limited dataset variety as areas that future research needs to address.