Towards Real-World Burst Image Super-Resolution: Benchmark and Method

This is the official implementation of ICCV2023 "Towards Real-World Burst Image Super-Resolution: Benchmark and Method".

Here is the [baidu link] (include Cropped RealBSR-RGB), the [baidu link] (include Full-size RealBSR-RGB) and the [google drive link] (include Cropped RealBSR-RGB and Cropped RealBSR-RAW) of the proposed RealBSR dataset. (Please read the folder structure below before using them)

Folder Structure of Data in Baidu Link of Cropped RealBSR-RGB

RealBSR Dataset
├── train
├──── LR (original LR. Need to run homography_alignment.py to align them.)
├──── LR_aligned (Aligned LR using homography. You can use the aligned data directly as inputs of network)
├──── GT (Corresponding HR)
├── test
├──── LR (original LR. Need to run homography_alignment.py to align them.)
├──── LR_aligned (Aligned LR using homography. You can use the aligned data directly as inputs of network)
├──── GT (Corresponding HR)

Folder Structure of Data in Baidu Link of Full-Size RealBSR-RGB

RealBSR Dataset
├── train
├──── LR (original LR. Need to run homography_alignment.py to align them.)
├──── rewarp.png (GT)
├── test
├──── LR (original LR. Need to run homography_alignment.py to align them.)
├──── rewarp.png (GT)

Folder Structure Of Data in Google Drive Link of Cropped RGB and RAW

RealBSR Dataset
├── RGB
├──── train
├────── LR_aligned.zip (Aligned LR using homography. You can use the aligned data directly as inputs)
├────── trainpatch.zip (Each group contains 14 original LR + 1 HR named xxxrewarp.png. If you want to use the LR data here, you need to run homography_alignment.py before regarding them as inputs of our network)
├──── test
├────── LR_aligned.zip (Aligned LR using homography. You can use the aligned data directly as inputs)
├────── testpatch.zip (Each group contains 14 original LR + 1 HR named xxxrewarp.png. If you want to use the LR data here, you need to run homography_alignment.py before regarding them as inputs of our network)

├── RAW
├──── trainpatch.zip (Each group contains 14 original LR RAW + 1 HR named xxx_rgb.png + 1 pkl named xxx.pkl. The raw LR need to be aligned first by simply changing the homography_alignment for 4 channels. )
├──── testpatch.zip (Each group contains 14 original LR RAW + 1 HR named xxx_rgb.png + 1 pkl named xxx.pkl. The raw LR need to be aligned first by simply changing the homography_alignment for 4 channels. )

Package dependencies

The project is built with PyTorch 1.10.1, Python3.9, CUDA10.1. For package dependencies, you can install them by:

pip3 install -r requirements.txt

Pretrained model

Checkpoint of FBANet on RealBSR dataset [Google Drive] and [Baidu Drive Link]

Model	PSNR	SSIM	LPIPS
FBANet (FAF in paper)	31.012	0.898	0.102

Data preparation

Alignment

For training and testing data of Burst, first you should align the burst frames using homography_alignment.py

python homography_alignment.py

The aligned frames could be put into '../LR_aligned' and then you should use the aligned frames as the input of the following training and testing.

Evaluation

If you want to use your own data, the resolution of data should better >=160x160, and the setting could be changed in ManualDataset.py and test_in_any_resolution.py.

For convenient evaluation, you could use data [here] for evaluation. It is worth noting that the data is not included in training or testing of our BaseModel.

test for arbitrary resolution

python homography_alignment.py
python test_in_any_resolution.py --arch BaseModel --weights ./checkpoints/FAF/model_best.pth --save_images

Training

Base Model

To train BaseModel on RealBSR, we use 2 V100 GPUs and run for 200 epochs:

python3 ./train.py --arch BaseModel --batch_size 16 --gpu '0,1' --train_ps 160 --env 64_0523_MotionMFSR_FAF --embed_dim 64 --warmup

Contact

My email is yujingsun1999@gmail.com. Please feel free to contact me about anything related to the paper. And I'm also very glad to talk about any topic of Super-Resolution!

yjsunnn/FBANet