Overview / Usage

This work illustrates multi-class open set classification of human actions recognized from short video clips. This project can be applied to various fields like smart surveillance, medicine, military security etc. The project is still under development.

Methodology / Approach

1. Classification of video data items into more than two classes to achieve multi-class classification. 3D-CNN approach will be used for this step.
2. Unknown sample should be classified as new/unknown category to achieve open set classification. Softmax Thresholding technique will be considered for same.
3. Determine the accuracy of the prediction
4. Perform evaluation of system using any evaluation measure such as F-measure, openness, occlusion experiment etc.

Technologies Used

1. Though system in this project is designed for only three classes, it should be scale-able to more number of classes.
2. Anaconda version of Python 2.7/3.7 is required for coding multi-class open set classification
3. Latest version of CUDA installed on GPU is required for high performance at time of training and testing
4. Pytorch and TorchVision libraries are required for utilization of deep learning frameworks.