"MST++: Multi-stage Spectral-wise Transformer for Efficient Spectral Reconstruction" (CVPRW 2022) & (Winner of NTIRE 2022 Challenge on Spectral Reconstruction from RGB)

Related tags

Deep Learningimage-restorationspectral-superresolutionspectral-reconstructionhyperspectral-image-reconstruction
Overview

MST++: Multi-stage Spectral-wise Transformer for Efficient Spectral Reconstruction (CVPRW 2022)

winner arXiv zhihu mst visitors

Yuanhao Cai, Jing Lin, Zudi Lin, Haoqian Wang, Yulun Zhang, Hanspeter Pfister, Radu Timofte, Luc Van Gool

The first two authors contribute equally to this work

News

  • 2022.04.17 : Our paper has been accepted by CVPRW 2022, code and models have been released. 🚀
  • 2022.04.02 : We win the First place of NTIRE 2022 Challenge on Spectral Reconstruction from RGB. 🏆
480 nm 520 nm 580 nm 660 nm

Abstract: Existing leading methods for spectral reconstruction (SR) focus on designing deeper or wider convolutional neural networks (CNNs) to learn the end-to-end mapping from the RGB image to its hyperspectral image (HSI). These CNN-based methods achieve impressive restoration performance while showing limitations in capturing the long-range dependencies and self-similarity prior. To cope with this problem, we propose a novel Transformer-based method, Multi-stage Spectral-wise Transformer (MST++), for efficient spectral reconstruction. In particular, we employ Spectral-wise Multi-head Self-attention (S-MSA) that is based on the HSI spatially sparse while spectrally self-similar nature to compose the basic unit, Spectral-wise Attention Block (SAB). Then SABs build up Single-stage Spectral-wise Transformer (SST) that exploits a U-shaped structure to extract multi-resolution contextual information. Finally, our MST++, cascaded by several SSTs, progressively improves the reconstruction quality from coarse to fine. Comprehensive experiments show that our MST++ significantly outperforms other state-of-the-art methods. In the NTIRE 2022 Spectral Reconstruction Challenge, our approach won the First place.

Network Architecture

Illustration of MST

Our MST++ is mainly based on our work MST, which is accepted by CVPR 2022.

Comparison with State-of-the-art Methods

This repo is a baseline and toolbox containing 11 image restoration algorithms for Spectral Reconstruction.

We are going to enlarge our model zoo in the future.

Supported algorithms:

comparison_fig

Results on NTIRE 2022 HSI Dataset - Validation

Method Params (M) FLOPS (G) MRAE RMSE PSNR Model Zoo
HSCNN+ 4.65 304.45 0.3814 0.0588 26.36 Google Drive / Baidu Disk
HRNet 31.70 163.81 0.3476 0.0550 26.89 Google Drive / Baidu Disk
EDSR 2.42 158.32 0.3277 0.0437 28.29 Google Drive / Baidu Disk
AWAN 4.04 270.61 0.2500 0.0367 31.22 Google Drive / Baidu Disk
HDNet 2.66 173.81 0.2048 0.0317 32.13 Google Drive / Baidu Disk
HINet 5.21 31.04 0.2032 0.0303 32.51 Google Drive / Baidu Disk
MIRNet 3.75 42.95 0.1890 0.0274 33.29 Google Drive / Baidu Disk
Restormer 15.11 93.77 0.1833 0.0274 33.40 Google Drive / Baidu Disk
MPRNet 3.62 101.59 0.1817 0.0270 33.50 Google Drive / Baidu Disk
MST-L 2.45 32.07 0.1772 0.0256 33.90 Google Drive / Baidu Disk
MST++ 1.62 23.05 0.1645 0.0248 34.32 Google Drive / Baidu Disk

Our MST++ siginificantly outperforms other methods while requiring cheaper Params and FLOPS.

Note: access code for Baidu Disk is mst1.

1. Create Envirement:

  • Python 3 (Recommend to use Anaconda)

  • NVIDIA GPU + CUDA

  • Python packages:

    cd MST-plus-plus
pip install -r requirements.txt

2. Data Preparation:

  • Download training spectral images (Google Drive / Baidu Disk, code: mst1), training RGB images (Google Drive / Baidu Disk), validation spectral images (Google Drive / Baidu Disk), validation RGB images (Google Drive / Baidu Disk), and testing RGB images (Google Drive / Baidu Disk) from the competition website of NTIRE 2022 Spectral Reconstruction Challenge.

  • Place the training spectral images and validation spectral images to /MST-plus-plus/dataset/Train_Spec/.

  • Place the training RGB images and validation RGB images to /MST-plus-plus/dataset/Train_RGB/.

  • Place the testing RGB images to /MST-plus-plus/dataset/Test_RGB/.

  • Then this repo is collected as the following form:

    |--MST-plus-plus
    |--test_challenge_code
    |--test_develop_code
    |--train_code  
    |--dataset 
        |--Train_Spec
            |--ARAD_1K_0001.mat
            |--ARAD_1K_0002.mat
            ： 
            |--ARAD_1K_0950.mat
  	|--Train_RGB
            |--ARAD_1K_0001.jpg
            |--ARAD_1K_0002.jpg
            ： 
            |--ARAD_1K_0950.jpg
        |--Test_RGB
            |--ARAD_1K_0951.jpg
            |--ARAD_1K_0952.jpg
            ： 
            |--ARAD_1K_1000.jpg
        |--split_txt
            |--train_list.txt
            |--valid_list.txt

3. Evaluation on the Validation Set:

(1) Download the pretrained model zoo from (Google Drive / Baidu Disk, code: mst1) and place them to /MST-plus-plus/test_develop_code/model_zoo/.

(2) Run the following command to test the model on the validation RGB images.

cd /MST-plus-plus/test_develop_code/

# test MST++
python test.py --data_root ../dataset/  --method mst_plus_plus --pretrained_model_path ./model_zoo/mst_plus_plus.pth --outf ./exp/mst_plus_plus/  --gpu_id 0

# test MST-L
python test.py --data_root ../dataset/  --method mst --pretrained_model_path ./model_zoo/mst.pth --outf ./exp/mst/  --gpu_id 0

# test MIRNet
python test.py --data_root ../dataset/  --method mirnet --pretrained_model_path ./model_zoo/mirnet.pth --outf ./exp/mirnet/  --gpu_id 0

# test HINet
python test.py --data_root ../dataset/  --method hinet --pretrained_model_path ./model_zoo/hinet.pth --outf ./exp/hinet/  --gpu_id 0

# test MPRNet
python test.py --data_root ../dataset/  --method mprnet --pretrained_model_path ./model_zoo/mprnet.pth --outf ./exp/mprnet/  --gpu_id 0

# test Restormer
python test.py --data_root ../dataset/  --method restormer --pretrained_model_path ./model_zoo/restormer.pth --outf ./exp/restormer/  --gpu_id 0

# test EDSR
python test.py --data_root ../dataset/  --method edsr --pretrained_model_path ./model_zoo/edsr.pth --outf ./exp/edsr/  --gpu_id 0

# test HDNet
python test.py --data_root ../dataset/  --method hdnet --pretrained_model_path ./model_zoo/hdnet.pth --outf ./exp/hdnet/  --gpu_id 0

# test HRNet
python test.py --data_root ../dataset/  --method hrnet --pretrained_model_path ./model_zoo/hrnet.pth --outf ./exp/hrnet/  --gpu_id 0

# test HSCNN+
python test.py --data_root ../dataset/  --method hscnn_plus --pretrained_model_path ./model_zoo/hscnn_plus.pth --outf ./exp/hscnn_plus/  --gpu_id 0

# test AWAN
python test.py --data_root ../dataset/  --method awan --pretrained_model_path ./model_zoo/awan.pth --outf ./exp/awan/  --gpu_id 0

The results will be saved in /MST-plus-plus/test_develop_code/exp/ in the mat format and the evaluation metric (including MRAE,RMSE,PSNR) will be printed.

4. Evaluation on the Test Set:

(1) Download the pretrained model zoo from (Google Drive / Baidu Disk, code: mst1) and place them to /MST-plus-plus/test_challenge_code/model_zoo/.

(2) Run the following command to test the model on the testing RGB images.

cd /MST-plus-plus/test_challenge_code/

# test MST++
python test.py --data_root ../dataset/  --method mst_plus_plus --pretrained_model_path ./model_zoo/mst_plus_plus.pth --outf ./exp/mst_plus_plus/  --gpu_id 0

# test MST-L
python test.py --data_root ../dataset/  --method mst --pretrained_model_path ./model_zoo/mst.pth --outf ./exp/mst/  --gpu_id 0

# test MIRNet
python test.py --data_root ../dataset/  --method mirnet --pretrained_model_path ./model_zoo/mirnet.pth --outf ./exp/mirnet/  --gpu_id 0

# test HINet
python test.py --data_root ../dataset/  --method hinet --pretrained_model_path ./model_zoo/hinet.pth --outf ./exp/hinet/  --gpu_id 0

# test MPRNet
python test.py --data_root ../dataset/  --method mprnet --pretrained_model_path ./model_zoo/mprnet.pth --outf ./exp/mprnet/  --gpu_id 0

# test Restormer
python test.py --data_root ../dataset/  --method restormer --pretrained_model_path ./model_zoo/restormer.pth --outf ./exp/restormer/  --gpu_id 0

# test EDSR
python test.py --data_root ../dataset/  --method edsr --pretrained_model_path ./model_zoo/edsr.pth --outf ./exp/edsr/  --gpu_id 0

# test HDNet
python test.py --data_root ../dataset/  --method hdnet --pretrained_model_path ./model_zoo/hdnet.pth --outf ./exp/hdnet/  --gpu_id 0

# test HRNet
python test.py --data_root ../dataset/  --method hrnet --pretrained_model_path ./model_zoo/hrnet.pth --outf ./exp/hrnet/  --gpu_id 0

# test HSCNN+
python test.py --data_root ../dataset/  --method hscnn_plus --pretrained_model_path ./model_zoo/hscnn_plus.pth --outf ./exp/hscnn_plus/  --gpu_id 0

The results and submission.zip will be saved in /MST-plus-plus/test_challenge_code/exp/.

5. Training

To train a model, run

cd /MST-plus-plus/train_code/

# train MST++
python train.py --method mst_plus_plus  --batch_size 20 --end_epoch 300 --init_lr 4e-4 --outf ./exp/mst_plus_plus/ --data_root ../dataset/  --patch_size 128 --stride 8  --gpu_id 0

# train MST-L
python train.py --method mst  --batch_size 20 --end_epoch 300 --init_lr 4e-4 --outf ./exp/mst/ --data_root ../dataset/  --patch_size 128 --stride 8  --gpu_id 0

# train MIRNet
python train.py --method mirnet  --batch_size 20 --end_epoch 300 --init_lr 4e-4 --outf ./exp/mirnet/ --data_root ../dataset/  --patch_size 128 --stride 8  --gpu_id 0

# train HINet
python train.py --method hinet  --batch_size 20 --end_epoch 300 --init_lr 2e-4 --outf ./exp/hinet/ --data_root ../dataset/  --patch_size 128 --stride 8  --gpu_id 0

# train MPRNet
python train.py --method mprnet  --batch_size 20 --end_epoch 300 --init_lr 2e-4 --outf ./exp/mprnet/ --data_root ../dataset/  --patch_size 128 --stride 8  --gpu_id 0

# train Restormer
python train.py --method restormer  --batch_size 20 --end_epoch 300 --init_lr 2e-4 --outf ./exp/restormer/ --data_root ../dataset/  --patch_size 128 --stride 8  --gpu_id 0

# train EDSR
python train.py --method edsr  --batch_size 20 --end_epoch 300 --init_lr 1e-4 --outf ./exp/edsr/ --data_root ../dataset/  --patch_size 128 --stride 8  --gpu_id 0

# train HDNet
python train.py --method hdnet  --batch_size 20 --end_epoch 300 --init_lr 4e-4 --outf ./exp/hdnet/ --data_root ../dataset/  --patch_size 128 --stride 8  --gpu_id 0

# train HRNet
python train.py --method hrnet  --batch_size 20 --end_epoch 300 --init_lr 1e-4 --outf ./exp/hrnet/ --data_root ../dataset/  --patch_size 128 --stride 8  --gpu_id 0

# train HSCNN+
python train.py --method hscnn_plus  --batch_size 20 --end_epoch 300 --init_lr 2e-4 --outf ./exp/hscnn_plus/ --data_root ../dataset/  --patch_size 128 --stride 8  --gpu_id 0

# train AWAN
python train.py --method awan  --batch_size 20 --end_epoch 300 --init_lr 1e-4 --outf ./exp/awan/ --data_root ../dataset/  --patch_size 128 --stride 8  --gpu_id 0

The training log and models will be saved in /MST-plus-plus/train_code/exp/.

6. Prediction:

(1) Download the pretrained model zoo from (Google Drive / Baidu Disk, code: mst1) and place them to /MST-plus-plus/predict_code/model_zoo/.

(2) Run the following command to reconstruct your own RGB image.

cd /MST-plus-plus/predict_code/

# reconstruct by MST++
python test.py --rgb_path ./demo/ARAD_1K_0912.jpg  --method mst_plus_plus --pretrained_model_path ./model_zoo/mst_plus_plus.pth --outf ./exp/mst_plus_plus/  --gpu_id 0

# reconstruct by MST-L
python test.py --rgb_path ./demo/ARAD_1K_0912.jpg  --method mst --pretrained_model_path ./model_zoo/mst.pth --outf ./exp/mst/  --gpu_id 0

# reconstruct by MIRNet
python test.py --rgb_path ./demo/ARAD_1K_0912.jpg  --method mirnet --pretrained_model_path ./model_zoo/mirnet.pth --outf ./exp/mirnet/  --gpu_id 0

# reconstruct by HINet
python test.py --rgb_path ./demo/ARAD_1K_0912.jpg  --method hinet --pretrained_model_path ./model_zoo/hinet.pth --outf ./exp/hinet/  --gpu_id 0

# reconstruct by MPRNet
python test.py --rgb_path ./demo/ARAD_1K_0912.jpg  --method mprnet --pretrained_model_path ./model_zoo/mprnet.pth --outf ./exp/mprnet/  --gpu_id 0

# reconstruct by Restormer
python test.py --rgb_path ./demo/ARAD_1K_0912.jpg  --method restormer --pretrained_model_path ./model_zoo/restormer.pth --outf ./exp/restormer/  --gpu_id 0

# reconstruct by EDSR
python test.py --rgb_path ./demo/ARAD_1K_0912.jpg --method edsr --pretrained_model_path ./model_zoo/edsr.pth --outf ./exp/edsr/  --gpu_id 0

# reconstruct by HDNet
python test.py --rgb_path ./demo/ARAD_1K_0912.jpg  --method hdnet --pretrained_model_path ./model_zoo/hdnet.pth --outf ./exp/hdnet/  --gpu_id 0

# reconstruct by HRNet
python test.py --rgb_path ./demo/ARAD_1K_0912.jpg  --method hrnet --pretrained_model_path ./model_zoo/hrnet.pth --outf ./exp/hrnet/  --gpu_id 0

# reconstruct by HSCNN+
python test.py --rgb_path ./demo/ARAD_1K_0912.jpg  --method hscnn_plus --pretrained_model_path ./model_zoo/hscnn_plus.pth --outf ./exp/hscnn_plus/  --gpu_id 0

You can replace './demo/ARAD_1K_0912.jpg' with your RGB image path. The reconstructed results will be saved in /MST-plus-plus/predict_code/exp/.

Citation

If this repo helps you, please consider citing our works:

@inproceedings{mst,
	title={Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction},
	author={Yuanhao Cai and Jing Lin and Xiaowan Hu and Haoqian Wang and Xin Yuan and Yulun Zhang and Radu Timofte and Luc Van Gool},
	booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
	year={2022}
}

@inproceedings{mst_pp,
  title={MST++: Multi-stage Spectral-wise Transformer for Efficient Spectral Reconstruction},
  author={Yuanhao Cai and Jing Lin and Zudi Lin and Haoqian Wang and Yulun Zhang and Hanspeter Pfister and Radu Timofte and Luc Van Gool},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  year={2022}
}

@inproceedings{hdnet,
	title={HDNet: High-resolution Dual-domain Learning for Spectral Compressive Imaging},
	author={Xiaowan Hu and Yuanhao Cai and Jing Lin and  Haoqian Wang and Xin Yuan and Yulun Zhang and Radu Timofte and Luc Van Gool},
	booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
	year={2022}
}
Owner
Yuanhao Cai
Tsinghua University [email protected]
Yuanhao Cai
GitHub Repository
Language models are open knowledge graphs ( non official implementation )

language-models-are-knowledge-graphs-pytorch Language models are open knowledge graphs ( work in progress ) A non official reimplementation of Languag

theblackcat102 132 Dec 18, 2022
Official Keras Implementation for UNet++ in IEEE Transactions on Medical Imaging and DLMIA 2018

UNet++: A Nested U-Net Architecture for Medical Image Segmentation UNet++ is a new general purpose image segmentation architecture for more accurate i

Zongwei Zhou 1.8k Dec 27, 2022
Vector Quantization, in Pytorch

Vector Quantization - Pytorch A vector quantization library originally transcribed from Deepmind's tensorflow implementation, made conveniently into a

Phil Wang 665 Jan 08, 2023
A series of convenience functions to make basic image processing operations such as translation, rotation, resizing, skeletonization, and displaying Matplotlib images easier with OpenCV and Python.

imutils A series of convenience functions to make basic image processing functions such as translation, rotation, resizing, skeletonization, and displ

Adrian Rosebrock 4.3k Jan 08, 2023
Jiminy Cricket Environment (NeurIPS 2021)

Jiminy Cricket This is the repository for "What Would Jiminy Cricket Do? Towards Agents That Behave Morally" by Dan Hendrycks*, Mantas Mazeika*, Andy

Dan Hendrycks 15 Aug 29, 2022
Implementation of Kronecker Attention in Pytorch

Kronecker Attention Pytorch Implementation of Kronecker Attention in Pytorch. Results look less than stellar, but if someone found some context where

Phil Wang 16 May 06, 2022
P-Tuning v2: Prompt Tuning Can Be Comparable to Finetuning Universally Across Scales and Tasks

P-tuning v2 P-Tuning v2: Prompt Tuning Can Be Comparable to Finetuning Universally Across Scales and Tasks An optimized prompt tuning strategy achievi

THUDM 540 Dec 30, 2022
A tensorflow implementation of GCN-LPA

GCN-LPA This repository is the implementation of GCN-LPA (arXiv): Unifying Graph Convolutional Neural Networks and Label Propagation Hongwei Wang, Jur

Hongwei Wang 83 Nov 28, 2022
Official page of Patchwork (RA-L'21 w/ IROS'21)

Patchwork Official page of "Patchwork: Concentric Zone-based Region-wise Ground Segmentation with Ground Likelihood Estimation Using a 3D LiDAR Sensor

Hyungtae Lim 254 Jan 05, 2023
Robust Lane Detection via Expanded Self Attention (WACV 2022)

Robust Lane Detection via Expanded Self Attention (WACV 2022) Minhyeok Lee, Junhyeop Lee, Dogyoon Lee, Woojin Kim, Sangwon Hwang, Sangyoun Lee Overvie

Min Hyeok Lee 18 Nov 12, 2022
Pre-trained model, code, and materials from the paper "Impact of Adversarial Examples on Deep Learning Models for Biomedical Image Segmentation" (MICCAI 2019).

Adaptive Segmentation Mask Attack This repository contains the implementation of the Adaptive Segmentation Mask Attack (ASMA), a targeted adversarial

Utku Ozbulak 53 Jul 04, 2022
This repository contains all the code and materials distributed in the 2021 Q-Programming Summer of Qode.

Q-Programming Summer of Qode This repository contains all the code and materials distributed in the Q-Programming Summer of Qode. If you want to creat

Sammarth Kumar 11 Jun 11, 2021
SIR model parameter estimation using a novel algorithm for differentiated uniformization.

TenSIR Parameter estimation on epidemic data under the SIR model using a novel algorithm for differentiated uniformization of Markov transition rate m

The Spang Lab 4 Nov 30, 2022
LaBERT - A length-controllable and non-autoregressive image captioning model.

Length-Controllable Image Captioning (ECCV2020) This repo provides the implemetation of the paper Length-Controllable Image Captioning. Install conda

bearcatt 53 Nov 13, 2022
Code of paper "Compositionally Generalizable 3D Structure Prediction"

Compositionally Generalizable 3D Structure Prediction In this work, We bring in the concept of compositional generalizability and factorizes the 3D sh

Songfang Han 30 Dec 17, 2022
Personal project about genus-0 meshes, spherical harmonics and a cow

How to transform a cow into spherical harmonics ? Spot the cow, from Keenan Crane's blog Context In the field of Deep Learning, training on images or

3 Aug 22, 2022
Reimplementation of the paper "Attention, Learn to Solve Routing Problems!" in jax/flax.

JAX + Attention Learn To Solve Routing Problems Reinplementation of the paper Attention, Learn to Solve Routing Problems! using Jax and Flax. Fully su

Gabriela Surita 7 Dec 01, 2022
This is the official implementation of our proposed SwinMR

SwinMR This is the official implementation of our proposed SwinMR: Swin Transformer for Fast MRI Please cite: @article{huang2022swin, title={Swi

A Yang Lab (led by Dr Guang Yang) 27 Nov 17, 2022
使用深度学习框架提取视频硬字幕；docker容器免安装深度学习库，使用本地api接口使得界面和后端识别分离；

extract-video-subtittle 使用深度学习框架提取视频硬字幕； 本地识别无需联网； CPU识别速度可观； 容器提供API接口； 运行环境 本项目运行环境非常好搭建，我做好了docker容器免安装各种深度学习包； 提供windows界面操作； 容器为CPU版本； 视频演示 https

歌者 16 Aug 06, 2022
Python PID Tuner - Based on a FOPDT model obtained using a Open Loop Process Reaction Curve

PythonPID_Tuner Step 1: Takes a Process Reaction Curve in csv format - assumes data at 100ms interval (column names CV and PV) Step 2: Makes a rough e

6 Jan 14, 2022