Graph(Graph): A Nested Graph-Based Framework for Early Accident Anticipation (WACV 2024)

Official PyTorch implementation of WACV 2024 paper - Graph(Graph): A Nested Graph-Based Framework for Early Accident Anticipation

Setup

• Python 3.9
• CUDA - 11.8

Create a conda environment and install all the dependencies using the following commands:

pip install -r requirements.txt

Dataset

For DAD and CCD:
Download the data from link and place it in data folder. There are 3 folders and 1 file for each dataset:

• obj_feat: The object data for both datasets is downloaded from [1].
• i3d_feat: We extracted I3D features for all the frames using the code and pretrained model available at [2].
• frames_stat: This contains the resolution for every frame of a video.
• obj_idx_to_labels.json: This contains classnames for object detections done in the feature extraction process.

For DoTA:
You can use the scripts in data/scripts/dota to generate dataset for the algorithm. (Here, vgg16 features are used instead of i3d features).There must be folowing 4 folders and 1 file for the dataset:

• obj_feat: The object data for both datasets is downloaded from [1].
• i3d_feat: We extracted I3D features for all the frames using the code and pretrained model available at [2].
• frames_stat: This contains the resolution for every frame of a video.
• toas: frame where accidents happen for all positive videos in text files
• obj_idx_to_labels.json: This contains classnames for object detections done in the feature extraction process.

Training

To train use the following commands: DAD dataset-

python train_dad.py --test_only 0  

DoTA dataset-

python train_dota.py --test_only 0  

CCD dataset-

python train_ccd.py --test_only 0 

The models will be saved in the model_checkpoints/'dataset-name' folder.

Evaluation

Download original trained models for DAD and CCD and trained models for DoTA from here. Place them in model_checkpoints folder.

Use the following command for evaluation:

DAD dataset-

python train_dad.py --test_only 1 --checkpoint_model "model_checkpoints/dad_model.pth" 

DoTA dataset-

python train_dota.py --test_only 1 --checkpoint_model "model_checkpoints/dota_model.pth" 

CCD dataset-

python train_ccd.py --test_only 1 --checkpoint_model "model_checkpoints/ccd_model.pth"

Cross Validation

You can perform cross validation on DoTA Dataset-

python cross_validate_dota.py --n_folds 5

References

1. https://github.com/Cogito2012/UString
2. https://github.com/piergiaj/pytorch-i3d
3. https://github.com/eriklindernoren/Action-Recognition
4. https://github.com/thakurnupur/Graph-Graph

Citation

please cite the original paper:

@inproceedings{thakur2024graph,
  title={Graph (Graph): A Nested Graph-Based Framework for Early Accident Anticipation},
  author={Thakur, Nupur and Gouripeddi, PrasanthSai and Li, Baoxin},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={7533--7541},
  year={2024}
}

In case of any questions, feel free to reach out at Kumudu Mohottala or open issues on the repo.