Skip to content

lihaoliu-cambridge/scotch-and-soda

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SCOTCH and SODA: A Transformer Video Shadow Detection Framework

This repository provides a Pytorch implementation of the paper "SCOTCH and SODA: A Transformer Video Shadow Detection Framework, CVPR'23".

Requirement

cuda==11.1
cudnn==8.0
torch==1.9.0
timm==0.9.6
transformers==4.30.2
pytorch-lightning==1.5.10
medpy==0.4.0
einops==0.6.1

Usage Instructions

  1. Setup

    Clone the repository and navigate to its directory:

    git clone https://github.com/lihaoliu-cambridge/scotch-and-soda.git
    cd scotch-and-soda
  2. Dataset Preparation

    Download and unzip ViSha dataset, Place the unzipped Visha directory into the dataset directory:

    ./dataset/ViSha
  3. Configuration

    Adjust the configurations for the dataloader, model architecture, and training logic in:

    ./config/scotch_and_soda_visha_image_config.yaml
  4. Training

    To train the model, execute:

    python train.py

    Note: Due to the large GPU memory requirement from the video-level dataloader, the dataloader has been switched to an image-level dataloader for easy training, which gives comparable results to the video-level dataloader. It's also advised to first train with the image-level dataloader and subsequently fine-tune with the video-level dataloader for fast convergency.

  5. Monitoring with Tensorboard

    To view the training progress, start Tensorboard and open http://127.0.0.1:6006/ in your browser:

    tensorboard --port=6006  --logdir=[Your Project Directory]/output/tensorboard/scotch_and_soda_visha_image
  6. Testing

    After training, update the checkpoint file path in the test.py script. Then, test the trained model using:

    python test.py

Results from Scotch and Soda

We have evaluated our "Scotch and Soda" model on the ViSha testing set. The results have been made available for viewing and download on Google Drive.

Citation

If you use this code or the associated paper in your work, please cite:

@inproceedings{liu2023scotch,
   title={SCOTCH and SODA: A Transformer Video Shadow Detection Framework},
   author={Liu, Lihao and Prost, Jean and Zhu, Lei and Papadakis, Nicolas and Li{\`o}, Pietro and Sch{\"o}nlieb, Carola-Bibiane and Aviles-Rivero, Angelica I},
   booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
   pages={10449--10458},
   year={2023}
}

About

[CVPR'23] Code for "SCOTCH and SODA: A Transformer Video Shadow Detection Framework".

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published