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}
}