This repository contains the code and resources used in the research paper "Comprehensive Analysis of the Jester Dataset Using State-of-the-Art Video Classification Models".
The paper thoroughly details our findings. Although it has not been formally published, access can be arranged by contacting one of the contributors to this repository.
The repository includes the following:
-
Dataset Preprocessing:
- Code used to generate training, validation, and test splits from the dataset;
- A script for calculating statistics on our splits;
- A script for calculating the mean and standard deviation of the Jester dataset frames;
-
Dataset Splits:
- The generated training, validation, and test splits, provided as space-separated text files;
-
Model Training and Testing:
- Code for training and testing state-of-the-art video classification models on the Jester dataset;
- A script for plotting the GFLOPS (inference) against the accuracy of the models we trained.
Due to the size of an average checkpoint, these have not been included in this repository. However, the checkpoints for the models we trained are available under the GPLv3 and can be accessed at https://files.jortvanleenen.nl/Jester-dataset-analysis.
The models in this repository are built using PyTorchVideo, which is partly based on and compatible with PyTorch and specifically torchvision. In addition, we use various packages for data handling, image processing, and numerical operations, such as Pillow, NumPy and tqdm.
This repository is licensed under the GNU General Public License v3.0 (GPLv3).