The project is the official implementation of our CVPR 2021 paper, "Restoring Extremely Dark Images in Real Time"
— Mohit Lamba, Kaushik Mitra
A practical low-light enhancement solution must be computationally fast, memory-efficient, and achieve a visually appealing restoration. Most of the existing methods target restoration quality and thus compromise on speed and memory requirements, raising concerns about their real-world deployability. We propose a new deep learning architecture for extreme low-light single image restoration, which despite its fast & lightweight inference, produces a restoration that is perceptually at par with state-of-the-art computationally intense models. To achieve this, we do most of the processing in the higher scale-spaces, skipping the intermediate-scales wherever possible. Also unique to our model is the potential to process all the scale-spaces concurrently, offering an additional 30% speedup without compromising the restoration quality. Pre-amplification of the dark raw-image is an important step in extreme low-light image enhancement. Most of the existing state of the art methods need GT exposure value to estimate the pre-amplification factor, which is not practically feasible. Thus, we propose an amplifier module that estimates the amplification factor using only the input raw image and can be used “off-the-shelf” with pre-trained models without any fine-tuning. We show that our model can restore an ultra-high-definition 4K resolution image in just 1 sec. on a CPU and at 32 fps on a GPU and yet maintain a competitive restoration quality. We also show that our proposed model, without any fine-tuning, generalizes well to cameras not seen during training and to subsequent tasks such as object detection.
The code was tested on Ubuntu 16.04 LTS and Google Colab with PyTorch 1.4. Apart from commonly used libraries you need to install rawpy pip install rawpy. To measure floating point operations you additionally require pip install ptflops.
To execute the following demos in Google Colab add %%bash at the beginning.
pip install rawpy
git clone https://github.com/MohitLamba94/Restoring-Extremely-Dark-Images-In-Real-Time.git
cd Restoring-Extremely-Dark-Images-In-Real-Time
python demo.py
The above code will read extreme low-light images present in demo_imgs, create demo_restored_images directory and save the restored images for different estimated amplification factor in this directory. The expected output in Google Colab is,
...... Loading all files to CPU RAM
Image No.: 1, Amplification_m=1: 22.907602310180664
Image No.: 2, Amplification_m=1: 53.080570220947266
Image No.: 3, Amplification_m=1: 45.878238677978516
Files loaded to CPU RAM......
Network parameters : 784768
Device on GPU: False
Restored images saved in DEMO_RESTORED_IMAGES directory
The restored images should look like this.
pip install ptflops
git clone https://github.com/MohitLamba94/Restoring-Extremely-Dark-Images-In-Real-Time.git
cd Restoring-Extremely-Dark-Images-In-Real-Time
python time_complexity.py
The expected output on google Colab is,
---Our Model parameters : 784768
---SID model parameters : 7760748
Computational complexity of Our model: 41.38 GMac
Computational complexity of SID model: 440.46 GMac
Beginning Warmup...
Time taken by our model on CPU for 8MP image : 1.827214765548706 seconds
Time taken by SID model on CPU for 8MP image : 9.998942331314088 seconds
We trained the network for 1 Million iterations on RAW images. As RAW images are bulky, reading them from storage is an expensive step. We therefore recommend loading the dataset into CPU RAM for 8-10x faster training.
If loading dataset into RAM first clear your swap,
swapoff -a && swapon -a
Then till the dataset loads into RAM, you may need to clear cache at regular intervals using,
sudo sh -c "/bin/echo 1 > /proc/sys/vm/drop_caches"
We provide a dummy train.py to give an overview of training and testing procedure and provide detailed comments to modify the code to your requiremnts.
@inproceedings{RealTimeDarkImageRestorationCvpr2021,
title={Restoring Extremely Dark Images in Real Time},
author={Lamba, Mohit and Mitra, Kaushik},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={3487--3497},
year={2021}
}
Copyright © Mohit Lamba, 2021. Patent Pending. All rights reserved. Please see the license file for terms.