Official repository for "Restormer: Efficient Transformer for High-Resolution Image Restoration". SOTA results for single-image motion deblurring, image deraining, image denoising (synthetic and real data), and dual-pixel defocus deblurring.

Overview

Restormer: Efficient Transformer for High-Resolution Image Restoration

Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang

Paper: https://arxiv.org/abs/2111.09881

Testing codes and pre-trained models will be released on Nov 22nd!


Abstract: Since convolutional neural networks (CNNs) perform well at learning generalizable image priors from large-scale data, these models have been extensively applied to image restoration and related tasks. Recently, another class of neural architectures, Transformers, have shown significant performance gains on natural language and high-level vision tasks. While the Transformer model mitigates the shortcomings of CNNs (i.e., limited receptive field and inadaptability to input content), its computational complexity grows quadratically with the spatial resolution, therefore making it infeasible to apply to most image restoration tasks involving high-resolution images. In this work, we propose an efficient Transformer model by making several key designs in the building blocks (multi-head attention and feed-forward network) such that it can capture long-range pixel interactions, while still remaining applicable to large images. Our model, named Restoration Transformer (Restormer), achieves state-of-the-art results on several image restoration tasks, including image deraining, single-image motion deblurring, defocus deblurring (single-image and dual-pixel data), and image denoising (Gaussian grayscale/color denoising, and real image denoising).


Network Architecture

Results

Image Deraining comparisons on the Test100, Rain100H, Rain100L, Test1200, and Test2800 testsets.


Single-Image Motion Deblurring results. Our Restormer is trained only on the GoPro dataset and directly applied to the HIDE and RealBlur benchmark datasets.


Defocus Deblurring comparisons on the DPDD testset (containing 37 indoor and 39 outdoor scenes). S: single-image defocus deblurring. D: dual-pixel defocus deblurring.


Gaussian Image Denoising comparisons for two categories of methods. Top super row: learning a single model to handle various noise levels. Bottom super row: training a separate model for each noise level.

Grayscale

Color

Real Image Denoising on SIDD and DND datasets. ∗ denotes methods using additional training data. Our Restormer is trained only on the SIDD images and directly tested on DND.

Citation

If you use Restormer, please consider citing:

@article{Zamir2021Restormer,
    title={Restormer: Efficient Transformer for High-Resolution Image Restoration}, 
    author={Syed Waqas Zamir and Aditya Arora and Salman Khan and Munawar Hayat 
            and Fahad Shahbaz Khan and Ming-Hsuan Yang},
    journal={ArXiv 2111.09881},
    year={2021}
}

Contact

Should you have any question, please contact [email protected]

Comments
  • Problems about training Deraining

    Problems about training Deraining

    Hi,Congratulations to you have a good job! Although I haved changed the number of GPUs in train.sh and Deraining_Restormer.yml to 4 since I only have 4 GPUs,I can't train the code of Deraining due to my GPU memory limitations. I found the program can run if I change the batch_size_per_gpu smaller. But the batch size can't meet the experimental settings. So what can I do if I want to achieve the settings in your experiment ( i.e. For progressive learning, we start training with patch size 128×128 and batch size 64. The patch size and batch size pairs are updated to [(160^2,40), (192^2,32), (256^2,16), (320^2,8),(384^2,8)] at iterations [92K, 156K, 204K, 240K, 276K].) ?

    opened by Lucky0775 5
  • colab?

    colab?

    I am pleased with your work; the level of completeness is really professional! Do you guys have any plan to release the code for Google Colab? Unfortunately, I can't run the code on my local machine due to some poor factors.

    opened by osushilover 5
  • Questions about the quantitative results of other methods?

    Questions about the quantitative results of other methods?

    Hi, How are the quantitative results calculated for the other methods in Restormer Table 1? Are you quoting their results directly or are you retraining them?

    Looking forward to your reply. Thank you!

    opened by C-water 3
  • Typical GPU memory requirements for training?

    Typical GPU memory requirements for training?

    I was trying to run training Restormer, and succeed to run it with 128x128 size.

    However my GPU memory runs out when trying to train the network with 256x256 size and a batch size larger than 2. My GPU is RTX3080 with 10GB memory.

    Do you know how much memory we need to train it on 256x256 size patch and batch size >= 8 ?

    opened by wonwoolee 3
  • Motion Debluring Train

    Motion Debluring Train

    Hi.Thank you so much for your open source work. When I trained motion_deblur, I found that the effect in the paper could not be achieved.

    1. I followed the dependency tutorial mentioned in the repository ,downloaded the gopro dataset, and used the provided crop method to prepare the training set and validation set.
    2. And use the Deblurring_Restormer.yml configuration file for training. In the configuration file I modified to use single GPU training.
    3. In another experiment, I modified the training strategy to fix the crop size to 128. But the results of both experiments were less than 31db, which was much lower than the results in the paper. I wonder if details are missing and why the results are so different.
    opened by niehen6174 3
  • About the training

    About the training

    How to solve the error of create_dataloader, create_dataset in init.py in the train.py file? Also what is the difference between training on basicsr documents and training on specific tasks (e.g. Deraining)?

    opened by SunYJLU 3
  • problem on the step ”Install gdrive using“

    problem on the step ”Install gdrive using“

    Dear author,I met a problem when input the code "go get github.com/prasmussen/gdrive"

    package golang.org/x/oauth2/google: unrecognized import path "golang.org/x/oauth2/google" (https fetch: Get https://golang.org/x/oauth2/google?go-get=1: dial tcp 172.217.163.49:443: i/o timeout)

    I want to know how to solve this.THANKS!

    opened by ZYQii 3
  • add model to Huggingface

    add model to Huggingface

    Hi, would you be interested in adding Restormer to Hugging Face Hub? The Hub offers free hosting, and it would make your work more accessible and visible to the rest of the ML community. We can setup an organization or a user account under which restormer can be added similar to github.

    Example from other organizations: Keras: https://huggingface.co/keras-io Microsoft: https://huggingface.co/microsoft Facebook: https://huggingface.co/facebook

    Example spaces with repos: github: https://github.com/salesforce/BLIP Spaces: https://huggingface.co/spaces/akhaliq/BLIP

    github: https://github.com/facebookresearch/omnivore Spaces: https://huggingface.co/spaces/akhaliq/omnivore

    and here are guides for adding spaces/models/datasets to your org

    How to add a Space: https://huggingface.co/blog/gradio-spaces how to add models: https://huggingface.co/docs/hub/adding-a-model uploading a dataset: https://huggingface.co/docs/datasets/upload_dataset.html

    Please let us know if you would be interested and if you have any questions, we can also help with the technical implementation.

    opened by AK391 3
  •  denoising training dataset

    denoising training dataset

    well done ! But can you tell me about your denoising-working ,what dataset your used? real training dataset and Gaussian Denoising dataset. Thank you very much!

    opened by 17346604401 3
  • some problems

    some problems

    Since no training code is given, I write my own training program to train Restormer. However, at the beginning of the training, I could only set batchsize to 48 due to the limitation of GPUs memory. However, I found that the loss would hardly decrease when the first 10,000 to 20,000 iteration was carried out, which verified that the PSNR remained unchanged at about 26.2. Is the training relatively slow, or what is the problem? And if prob, I would like to know the upward trend of Val PSNR and the downward trend of loss during your training

    opened by jiaaihhy 3
  • Would you inform about the wide-shallow network?

    Would you inform about the wide-shallow network?

    Hello,

    In the ablation study, you compared deeper vs wider Restormer. I'm wondering about the wider Restormer you mentioned, so could you inform me of the details of it?

    opened by amoeba04 2
  • About lr_scheduler.py

    About lr_scheduler.py

    Hi ! In lr_scheduler.py, from torch.optim.lr_scheduler import _LRScheduler The message Cannot find reference '_LRScheduler' in 'lr_scheduler.pyi' How can I solve this problem?

    opened by Spacei567 3
  • Question about training denoising model

    Question about training denoising model

    I followed the instructions and conducted 2 Gaussian color image denoising experiments where sigma=15 and 50. But I can't reproduce the same PSNR value as paper shows. Here are my results: sigma=15 For CBSD68 dataset Noise Level 15 PSNR: 34.398237 For Kodak dataset Noise Level 15 PSNR: 35.439437 For McMaster dataset Noise Level 15 PSNR: 35.556497 For Urban100 dataset Noise Level 15 PSNR: 35.058984 sigma=50 For CBSD68 dataset Noise Level 50 PSNR: 28.586302 For Kodak dataset Noise Level 50 PSNR: 29.967525 For McMaster dataset Noise Level 50 PSNR: 30.237451 For Urban100 dataset Noise Level 50 PSNR: 29.891585

    Did I miss some important details?

    opened by Andrew0613 0
  • About PSNR of IFAN in defocus deblurring tasks (DPDD datasets).

    About PSNR of IFAN in defocus deblurring tasks (DPDD datasets).

    Hi, Did you retrain the IFAN on DPDD? IFAN only provided the results from 8bit images, which is inconsistent with the results in this paper.
    I guess you have retrained IFAN. If convenient, could you please provide the test pictures?

    Thank you very much!

    opened by C-water 0
  • About training, NCCL

    About training, NCCL

    RuntimeError: NCCL error in: /opt/conda/conda-bld/pytorch_1616554786529/work/torch/lib/c10d/ProcessGroupNCCL.cpp:33, unhandled cuda error, NCCL version 2.7.8 ncclUnhandledCudaError: Call to CUDA function failed.

    How can i fix it????? Plz help!

    opened by jjjjzyyyyyy 1
  • About training

    About training

    Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.

    opened by jjjjzyyyyyy 0
  • About the Gaussian color image denoising results.

    About the Gaussian color image denoising results.

    Hi, there is a question about the Gaussian color image denoising results on the Kodak24 dataset. I have downloaded the provided pre-trained models and use them for testing, under the provided code base and environment. However, I can not get the similar results on Kodak24 as you have reported in Table 5 of the main paper. In fact, I get lower PSNR values of testing on Kodak24 (e,g,. -0.12 dB for sigma15, -0.11 dB of sigma25, -0.14 dB of sigma 50). Can you give some explanations or suggestions? Thanks very much.

    opened by gladzhang 0
Owner
Syed Waqas Zamir
Research Scientist
Syed Waqas Zamir
Shuffle Attention for MobileNetV3

SA-MobileNetV3 Shuffle Attention for MobileNetV3 Train Run the following command for train model on your own dataset: python train.py --dataset mnist

Sajjad Aemmi 36 Dec 28, 2022
Model Quantization Benchmark

Introduction MQBench is an open-source model quantization toolkit based on PyTorch fx. The envision of MQBench is to provide: SOTA Algorithms. With MQ

500 Jan 06, 2023
HyperSeg: Patch-wise Hypernetwork for Real-time Semantic Segmentation Official PyTorch Implementation

: We present a novel, real-time, semantic segmentation network in which the encoder both encodes and generates the parameters (weights) of the decoder. Furthermore, to allow maximal adaptivity, the w

Yuval Nirkin 182 Dec 14, 2022
Multimodal Temporal Context Network (MTCN)

Multimodal Temporal Context Network (MTCN) This repository implements the model proposed in the paper: Evangelos Kazakos, Jaesung Huh, Arsha Nagrani,

Evangelos Kazakos 13 Nov 24, 2022
This repository lets you interact with Lean through a REPL.

lean-gym This repository lets you interact with Lean through a REPL. See Formal Mathematics Statement Curriculum Learning for a presentation of lean-g

OpenAI 87 Dec 28, 2022
Bio-OFC gym implementation and Gym-Fly environment

Bio-OFC gym implementation and Gym-Fly environment This repository includes the gym compatible implementation of the Bio-OFC algorithm from the paper

Siavash Golkar 1 Nov 16, 2021
A high-performance Python-based I/O system for large (and small) deep learning problems, with strong support for PyTorch.

WebDataset WebDataset is a PyTorch Dataset (IterableDataset) implementation providing efficient access to datasets stored in POSIX tar archives and us

1.1k Jan 08, 2023
Run object detection model on the Raspberry Pi

Using TensorFlow Lite with Python is great for embedded devices based on Linux, such as Raspberry Pi.

Dimitri Yanovsky 6 Oct 08, 2022
Pytorch implementation for "Density-aware Chamfer Distance as a Comprehensive Metric for Point Cloud Completion" (NeurIPS 2021)

Density-aware Chamfer Distance This repository contains the official PyTorch implementation of our paper: Density-aware Chamfer Distance as a Comprehe

Tong WU 93 Dec 15, 2022
Wikidated : An Evolving Knowledge Graph Dataset of Wikidata’s Revision History

Wikidated Wikidated 1.0 is a dataset of Wikidata’s full revision history, which encodes changes between Wikidata revisions as sets of deletions and ad

Lukas Schmelzeisen 11 Aug 16, 2022
Code for "Steerable Pyramid Transform Enables Robust Left Ventricle Quantification"

Code for "Steerable Pyramid Transform Enables Robust Left Ventricle Quantification" This is an end-to-end framework for accurate and robust left ventr

2 Jul 09, 2022
Solution of Kaggle competition: Sartorius - Cell Instance Segmentation

Sartorius - Cell Instance Segmentation https://www.kaggle.com/c/sartorius-cell-instance-segmentation Environment setup Build docker image bash .dev_sc

68 Dec 09, 2022
VR Viewport Pose Model for Quantifying and Exploiting Frame Correlations

This repository contains the introduction to the collected VRViewportPose dataset and the code for the IEEE INFOCOM 2022 paper: "VR Viewport Pose Model for Quantifying and Exploiting Frame Correlatio

0 Aug 10, 2022
Official code for: A Probabilistic Hard Attention Model For Sequentially Observed Scenes

"A Probabilistic Hard Attention Model For Sequentially Observed Scenes" Authors: Samrudhdhi Rangrej, James Clark Accepted to: BMVC'21 A recurrent atte

5 Nov 19, 2022
LibMTL: A PyTorch Library for Multi-Task Learning

LibMTL LibMTL is an open-source library built on PyTorch for Multi-Task Learning (MTL). See the latest documentation for detailed introductions and AP

765 Jan 06, 2023
PyTorch implementation of paper: AdaAttN: Revisit Attention Mechanism in Arbitrary Neural Style Transfer, ICCV 2021.

AdaAttN: Revisit Attention Mechanism in Arbitrary Neural Style Transfer [Paper] [PyTorch Implementation] [Paddle Implementation] Overview This reposit

148 Dec 30, 2022
A PyTorch library and evaluation platform for end-to-end compression research

CompressAI CompressAI (compress-ay) is a PyTorch library and evaluation platform for end-to-end compression research. CompressAI currently provides: c

InterDigital 680 Jan 06, 2023
Learning based AI for playing multi-round Koi-Koi hanafuda card games. Have fun.

Koi-Koi AI Learning based AI for playing multi-round Koi-Koi hanafuda card games. Platform Python PyTorch PySimpleGUI (for the interface playing vs AI

Sanghai Guan 10 Nov 20, 2022
A Python training and inference implementation of Yolov5 helmet detection in Jetson Xavier nx and Jetson nano

yolov5-helmet-detection-python A Python implementation of Yolov5 to detect head or helmet in the wild in Jetson Xavier nx and Jetson nano. In Jetson X

12 Dec 05, 2022
The goal of the exercises below is to evaluate the candidate knowledge and problem solving expertise regarding the main development focuses for the iFood ML Platform team: MLOps and Feature Store development.

The goal of the exercises below is to evaluate the candidate knowledge and problem solving expertise regarding the main development focuses for the iFood ML Platform team: MLOps and Feature Store dev

George Rocha 0 Feb 03, 2022