VRFF

VRFF is a Video Registration and Fusion Framework. VRFF effectively alleviates the flickering issues in video stream fusion. The VRFF framework is sufficiently flexible to rapidly integrate the latest image registration and fusion methods, demonstrating enhanced performance when combined with the VRFF framework. The framework diagram of VRFF is as follows:

Note:

When new registration and fusion tasks need to be integrated, they can be encapsulated and extended in the model/reg and model/fuse folders.

TODO

• Fixed bug.
• Update the new reg and fuse modules.

Install

Download this project:

git clone [email protected]:Meng-Sang/VRFF.git
cd VRFF

Runtime environment

This project requires a runtime environment with Python=3.8 or higher.

conda create -n VRFF  python=3.8
pip install -r requirement.txt

Download weights

Download the weight of registration.

# If the path model/reg/MatchFormer/weights does not exist, please run the following commands:
# mkdir -p model/reg/MatchFormer/weights
cd model/reg/MatchFormer/weights
wget https://github.com/Meng-Sang/SA-DNet/blob/master/model/reg/MatchFormer/weights/model.ckpt

Download the weight of fusion.

# If the path model/fuse/MatchFormer/weights does not exist, please run the following commands:
# mkdir -p model/fuse/MatchFormer/weights
cd model/fuse/MatchFormer/weights
wget https://github.com/Meng-Sang/SA-DNet/blob/master/model/fuse/UFusion/weights/model.pth

Demo

Download image pairs

Please download the image pair sequences from MMVS and HDO, and save them to the specified image_pairs_root_path location. The internal structure of image_pairs_root_path is as follows:

image_pairs_root_path:
  -ir:
    1.jpg
    ...
  -vi:
    1.jpg
    ...

Run with Python

from utils.predict_video import predict_video

if __name__ == "__main__":
    predict_video(image_pairs_root_path, size=(320, 240), ipf_n=47, moment_alpha=0.99, save_path="assets/results/ipf_mom_fuse.avi",is_show=True)
    predict_video(image_pairs_root_path, size=(320, 240), ipf_n=1, moment_alpha=None, save_path="assets/results/fuse.avi", is_show=True)

Show result

"4FRnNpmSmwktFJKjg" video of the MMVS dataset
"a" video of the HDO dataset
Traditional method	VRFF

Related Works:

The related datasets and projects involved in this work are as follows:

• MMVS
• HDO
• SA-DNet

Citing

@inproceedings{sang2024vrff,
  title={VRFF: Video Registration and Fusion Framework},
  author={Meng Sang, Housheng Xie and Yang Yang},
  booktitle={Proceedings of 2024 International Joint Conference on Neural Networks (IJCNN)},
  year={2024},	
  organization={IEEE (Institute of Electrical and Electronics Engineers)}
}