esjo93 / NTIRE-2021-Dehazing-Challenge

CVPR 2021 NTIRE workshop project repository of team VIP_UNIST

NTIRE 2021 NonHomogeneous Dehazing Challenge: UNIST VIP Lab

Introduction

This is our project repository for CVPR 2021 workshop.

The repository is an official implementation of the paper Multi-Scale Selective Residual Learning for Non-Homogeneous Dehazing.

Network Architecture

Dataset Preparation

You can download NTIRE 2021 NonHomogeneous Dehazing Challenge dataset after participating the challenge in the following link: https://competitions.codalab.org/competitions/28032

Your dataset directory should be composed of three directories as following:

dataset_directory
|-- train
|   |-- HAZY
|   |   |-- 01
|   |   |-- 02
|   |   `-- ...
|   `-- GT
|       |-- 01
|       |-- 02
|       `-- ...
|-- val
|   |-- HAZY
|   |   `-- ...
|   `-- GT
|       `-- ...
`-- test
    `-- HAZY
        `-- ...

Train

You can start training your model by following command:

$ python main.py train
Additional arguments:
    --data-dir: Dataset directory
    --gpus: Invidual GPU number(s) to use(e.g. 0,1,2)
    --name: Name of the experiment
    --batch-size: Training batch size
    --epochs: The number of total epochs
    --lr: Initial learning rate
    --lr-mode: Learning rate scheduling mode
    --step: Step size for learning rate decay
    --weight-decay: Weight decay factor
    --crop-size: Random crop size for data augmentation
    --random-scale: Random scaling ratio for data augmentation
    --random-rotate: Random rotating degree for data augmentation
    --random-identity-mapping: Random identity mapping ratio for data augmentation

Test

You can test your pretrained model by following command:

$ python main.py test -d [data path] --resume [pretrained model path] --phase test --batch-size 1

Download pretrained model: [download]

Results

1. Quantitative results on test set (#31~35)

Metrics	Test Scores (#31~35)
PSNR	19.156
SSIM	0.809
LPIPS1 (Alex-net)	0.205
LPIPS2 (VGG16)	0.227

2. Qualitative results on both validation set and test set

Download qualitative results of the pretrained model: [download]

Citation

If you find this code useful for your publications, please consider citing.

@InProceedings{Jo_2021_CVPR,
    author    = {Jo, Eunsung and Sim, Jae-Young},
    title     = {Multi-Scale Selective Residual Learning for Non-Homogeneous Dehazing},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2021},
    pages     = {507-515}
}