SID-Paddle

English | 简体中文

• SID-Paddle
  • 1 Introduction
  • 2 Reproduction Accuracy
  • 3 Dataset
  • 4 Environment
  • 5 Quick Start
    • step1: Clone
    • step2: Download Data
    • step3: Train
    • step4: Test
    • step5: Evaluate
  • 6 Code Structure and Explanation
    • 6.1 Code Structure
    • 6.2 Parameter Explanation
  • 7 Information

1 Introduction

This is a Paddlepaddle implementation of：Learning to See in the Dark in CVPR 2018, by Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun.

Raw code：Learning-to-See-in-the-Dark

Paper：Learning to See in the Dark

This code includes the default model for training and testing on the See-in-the-Dark (SID) dataset.

2 Reproduction Accuracy

Index	Raw Paper	Raw Code	Reproduction
PSNR	28.88	28.96	28.82
SSIM	0.787	0.785	0.787

3 Dataset

The dataset is SID-Sony. If you only use Sony dataset, please download the files started with Sony. After you download all the parts, you can combine them together by running: "cat SonyPart* > Sony.zip" and "cat FujiPart* > Fuji.zip".

• As mentioned in raw README, the authors found some misalignment with the ground-truth for image 10034, 10045, 10172. Please remove those images for quantitative results, but they still can be used for qualitative evaluations.

• The dataset contains 2697 short-exposed images and 231 long-exposed images. Note that multiple short-exposed images may correspond to the same long-exposed image.

  • Train set：1865 short-exposed +161 long-exposed images
  • Validation set：234 short-exposed +20 long-exposed images
  • Test set：598 short-exposed +50 long-exposed images

• Data format: The proposed method is designed for sensor raw data. The pretrained model probably not work for data from another camera sensor. We do not have support for other camera data. It also does not work for images after camera ISP, i.e., the JPG or PNG data.

4 Environment

• Hardware: Paddle AI Studio 4 * Tesla V100, 128G. It takes 16 hours for 4000 epochs. The GPU memory should be greater than 64 GB.
• Frameworks:
  • Paddlepaddle >= 2.0.0
  • rawpy
  • scipy == 1.1.0

5 Quick Start

Step1: Clone

# clone this repo
git clone git://github.com/WangChen0902/SID-Paddle.git

After downloading, please create three folders: "data", "checkpoint" and "result" in the root path.

Step2: Download Data

Place the datafiles mentioned above in "data" directory as format: "SID-Paddle/data/Sony".

Place the pre-trained models in "checkpoint" directory as format: "SID-Paddle/checkpoint".

Step3: Train

python train_Sony_paddle.py  # Single GPU
python -m paddle.distributed.launch train_Sony_paddle.py  # Multi-GPU

Step4: Test

python test_Sony_paddle.py

Step5: Evaluate

python eval.py

6 Code Structure and Explanation

6.1 Code Structure

├── checkpoint  # path to saved models
├── data  # path to data
├── result  # path to output
├── utils  # utils
│   ├── PSNR.py
│   └── SSIM.py
├── eval.py  # the script of calculating PSNR/SSIM
├── README.md
├── run.sh  # the script AI Studio multi-GPU training
├── run_test.sh  # the script of single GPU testing
├── test_Sony_paddle.py  # Testing
└── train_Sony_paddle.py  # Training

6.2 Parameter Explanation

Parameter Name	Default Value	Description
start_epoch	0	start epoch
num_epoches	4001	number of epochs
patch_size	512	patch size of training images
save_freq	200	save frequency of training
learning_rate	1e-4	learning rate
DEBUG	0	if debug
data_prefix	'./data/'	path to data
output_prefix	'./result/'	path to output
checkpoint_load_dir	'./checkpoint/'	path to checkpoint
last_epoch	4000	epoch of testing

7 Model Information

Information Name	Description
Author	Wangchen0902
Time	2021.08
Framework	Paddle 2.1.2
Application Scenario	Image enhancement
Hardware	GPU>=64G
Download	Pretrained model Access code: 6hbx
Download	Training logs Access code: brfz
Online Running	SID-Paddle