Spectroformer: Multi-Domain Query Cascaded Transformer Network For Underwater Image Enhancement

Md Raqib Khan · Priyanka Mishra · Nancy Mehta · Shruti S. Phutke · Santosh Kumar Vipparthi · Sukumar Nandi · Subrahmanyam Murala

WACV-2024

Paper

Evaluation

To evaluate the model on different datasets using the provided checkpoints and sample degraded images.

Dataset and Checkpoint Structure

• Sample degraded images for testing: Available in dataset/dataset_name/.
• Checkpoints for evaluation: Provided in checkpoints/dataset-name/.
• Results storage: After successful execution, the results will be saved in the results/dataset-name/ folder.

Folder Overview

├── dataset
│   ├── UIEB
│   ├── U-45
│   ├── SQUID
│   ├── UCCS
├── checkpoints
│   ├── UIEB
│   ├── U-45
│   ├── SQUID
│   ├── UCCS
├── results
│   ├── UIEB
│   ├── U-45
│   ├── SQUID
│   ├── UCCS

Running the Evaluation

To evaluate the model on different datasets, follow the instructions below for each specific dataset:

UIEB Dataset Evaluation

Run the following command to evaluate the model on the UIEB dataset:

python test.py --dataset datasets/UIEB/ --save_path Results/UIEB

U-45 Dataset Evaluation

Run the following command to evaluate the model on the U-45 dataset:

python test.py --dataset dataset/U-45/ --save_path Results/U-45

SQUID Dataset Evaluation

Run the following command to evaluate the model on the SQUID dataset:

python test.py --dataset dataset/SQUID/ --save_path Results/SQUID

UCCS Dataset Evaluation

Run the following command to evaluate the model on the UCCS dataset:

python test.py --dataset dataset/UCCS/ --save_path Results/UCCS

Traing

1. Structure of data for training should be like

uw_data/
   ├── train/
   │   ├── a/  # Input images
   │   └── b/  # Reference (ground truth) images
   └── test/
       ├── a/  # Input images
       └── b/  # Reference (ground truth) images

3. run

  pyhthon train.py

Citation

If you find this work helpful, please reference it as follows:

@inproceedings{khan2024spectroformer,
  title={Spectroformer: A Multi-Domain Query Cascaded Transformer Network for Underwater Image Enhancement},
  author={Khan, Raqib and Mishra, Priyanka and Mehta, Nancy and Phutke, Shruti S and Vipparthi, Santosh Kumar and Nandi, Sukumar and Murala, Subrahmanyam},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={1454--1463},
  year={2024}}

Acknowledgements

Special thanks to the awesome repositories UIPTA and Restoremer, which made this project possible.