This repository provides the official PyTorch implementation of the following paper:

Rethinking Coarse-to-Fine Approach in Single Image Deblurring

Sung-Jin Cho *, Seo-Won Ji *, Jun-Pyo Hong, Seung-Won Jung, Sung-Jea Ko

In ICCV 2021. (* indicates equal contribution)

Paper: https://arxiv.org/abs/2108.05054

Abstract: Coarse-to-fine strategies have been extensively used for the architecture design of single image deblurring networks. Conventional methods typically stack sub-networks with multi-scale input images and gradually improve sharpness of images from the bottom sub-network to the top sub-network, yielding inevitably high computational costs. Toward a fast and accurate deblurring network design, we revisit the coarse-to-fine strategy and present a multi-input multi-output U-net (MIMO-UNet). The MIMO-UNet has three distinct features. First, the single encoder of the MIMO-UNet takes multi-scale input images to ease the difficulty of training. Second, the single decoder of the MIMO-UNet outputs multiple deblurred images with different scales to mimic multi-cascaded U-nets using a single U-shaped network. Last, asymmetric feature fusion is introduced to merge multi-scale features in an efficient manner. Extensive experiments on the GoPro and RealBlur datasets demonstrate that the proposed network outperforms the state-of-the-art methods in terms of both accuracy and computational complexity.

The contents of this repository are as follows:

1. Dependencies
2. Dataset
3. Train
4. Test
5. Performance
   • GPU syncronization issue on measuring inference time
6. Model

• Python
• Pytorch (1.4)
  • Different versions may cause some errors.
• scikit-image
• opencv-python
• Tensorboard

• Download deblur dataset from the GoPro dataset .

• Unzip files dataset folder.

• Preprocess dataset by running the command below:

  python data/preprocessing.py

After preparing data set, the data folder should be like the format below:

GOPRO
├─ train
│ ├─ blur    % 2103 image pairs
│ │ ├─ xxxx.png
│ │ ├─ ......
│ │
│ ├─ sharp
│ │ ├─ xxxx.png
│ │ ├─ ......
│
├─ test    % 1111 image pairs
│ ├─ ...... (same as train)

To train MIMO-UNet+ , run the command below:

python main.py --model_name "MIMO-UNetPlus" --mode "train" --data_dir "dataset/GOPRO"

or to train MIMO-UNet, run the command below:

python main.py --model_name "MIMO-UNet" --mode "train" --data_dir "dataset/GOPRO"

Model weights will be saved in results/model_name/weights folder.

To test MIMO-UNet+ , run the command below:

python main.py --model_name "MIMO-UNetPlus" --mode "test" --data_dir "dataset/GOPRO" --test_model "MIMO-UNetPlus.pkl"

or to test MIMO-UNet, run the command below:

python main.py --model_name "MIMO-UNet" --mode "test" --data_dir "dataset/GOPRO" --test_model "MIMO-UNet.pkl"

Output images will be saved in results/model_name/result_image folder.

We recently found an issue about measuring the inference time of networks implemented using the PyTorch framework.

The official codes of many papers (more than twenty papers at a glance) presented at the top conferences measured the inference time simply using time measuring functions such as time.time(), time.perf_counter(), or tqdm. However, since the CUDA calls are asynchronous, the synchronized inference time needs to be measured using torch.cuda.synchronize().

We thus present Table and Figure containing the re-measured inference time using the synchronization mode for various methods developed with the PyTorch framework as shown below.

The inference times presented below were all measured using an RTX3090 due to the recent upgrade of our system. (The use of VRAM was restricted to 12 GB, which is the same value as that of Titan XP)

* indicates inference time measured without torch.cuda.synchronize().

** indicates inference time measured with torch.cuda.synchronize().

*** In case for MIMO-UNet++, we used batch inference (inference two times with batch size of 2) for geometrical ensemble. However, we noticed that the inference time measured with torch.cuda.synchronize() cannot take advantage of the batch inference, resulting MIMO-UNet++ performed 4x slower than MIMO-UNet+. (still 32% faster than the conventional SOTA method)

As the GPU was changed from Titan Xp to 3090, the asynchronized inference times of the conventional methods were reduced, but the inference times of MIMO-UNet and its variants were maintained. We will conduct additional tests on this issue and will update this page if there is any progress.

We hope this observation will be helpful to many researchers in this field.

We provide our pre-trained models. You can test our network following the instruction above.

• MIMO-UNet: https://drive.google.com/file/d/1EQJoQj3YMLFfzrbgzWMD3Xj96RqLdIlx/view?usp=sharing
• MIMO-UNet+: https://drive.google.com/file/d/166sufeHcdDTgXHNbCRzTC4T6DzuflB5m/view?usp=sharing
• MIMO-UNet+ (RealBlur): https://drive.google.com/file/d/1N3n-ozw1twt96ohX4tjsSAIcHrlNwHZ7/view?usp=sharing

* The test code for RealBlur dataset will be released soon.

** We measured PSNR using official RealBlur test code. You can get the PSNR we achieved by cloning and following the RealBlur repository.