Action Recognition for UCF 101

Final project for my Computer Vision class, to recognize class of a given video based. So far limited work has been done in using temporal features for feature detection, my aim is to either compare the latest state of the art tools for action recognition or to implement a vanilla solution.

The dataset used here is UCF 101, you can read more about it on the website. The first train/test splits are used from the three train/test splits provided.

Workflow

Step one

• Load videos from training file trainlist01.txt) convert it to frames.
• Extracting two frames per second.
• Save it in a folder named UCF101_images, and create a file named image_train1.txt
• Run a CNN over these images with their labels.

Step two

• After training the CNN, pass the last layer of CNN to LSTM

Installation

This requires python2.7 and pytorch.

Install the dependencies and devDependencies and start the server.

cd action_recognition_ucf101
conda create --name action
source activate action
pip install scikit-video
conda install pytorch torchvision -c soumith
conda install jupyter