Nmegha2601 / activitygraph_transformer

Activity Graph Transformer for Temporal Action Localization

This repository contains the implementation of Activity Graph Transformers. The paper is available here or at the project website.

---

Model

See the following image for an overview of the architecture of Activity Graph Transformer.

---

Getting Started

1. Clone the respository.

https://github.com/Nmegha2601/activitygraph_transformer.git

2. Use Conda to install Python3. Create a new environment in conda, and install dependencies from requirements.txt.

conda create --name agt_env --file requirements.txt

3. Extract I3D features. We used this to extract the features for each of the dataset in .npy format. Place the extracted features in data/thumos/i3d_feats

4. Run the code using the following command. This will start training a model on the dataset. The checkpoints and training log will be saved in an automatically created directory output. To reproduce the results, use the hyperparameters mentioned in the Appendix section of the paper.

bash run_scripts/run_agt.sh

---

Contact

For further questions, please email Megha Nawhal.