This is an official implementation of the paper "Learning with Privileged Information for Efficient Image Super-Resolution", accepted to ECCV2020.
This work effectively boosts the performance of FSRCNN by exploiting a distillation framework, treating HR images as privileged information.
For more information, checkout the project site [website] and the paper [PDF].
- Python 3.6
- PyTorch >= 1.1.0
- Please check requirements.txt
We provide a Dockerfile to reproduce our work easily.
$ docker build -t pisr:latest . # or docker pull wonkyunglee/pytorch_pisr:latest
$ docker run -it -v <working_dir>:/data -w /data pisr:latest /bin/bash- For training and validation
- DIV2K
- For evaluation
- Set5
- Set14
- B100
- Urban100
Please download DIV2K dataset from here and other benchmark datasets from here.
After download all datasets, the folder data should be like this:
data
├── benchmark
│ ├── B100
│ ├── Set14
│ ├── Set5
│ └── Urban100
│ ├── HR
│ └── LR_bicubic
│ ├── X2
│ ├── X3
│ └── X4
│
└── DIV2K
├── DIV2K_train_HR
└── DIV2K_train_LR_bicubic
├── X2
├── X3
└── X4
First, clone our github repository.
$ git clone https://github.com/yonsei-cvlab/PISR.gitTo train our teacher model, run the following script.
$ python step1_train_teacher.py --config configs/fsrcnn/step1.ymlTo train our student model, run the following script.
$ python step2_train_student.py --config configs/fsrcnn/step2.yml- Download pre-trained weights for teacher model into
results/fsrcnn/fsrcnn_teacher/checkpoint/folder.
Link: [weights] - Download pre-trained weights for student model into
results/fsrcnn/fsrcnn_student/checkpoint/folder.
Link: [weights]
To evaluate our student model, run following script. Benchmark datasets can be choosed by editing the config file configs/fsrcnn/base.ram.yml.
$ python evaluate.py --config configs/fsrcnn/step2.yml
@inproceedings{lee2020pisr,
author={Lee, Wonkyung and Lee, Junghyup and Kim, Dohyung and Ham, Bumsub},
title={Learning with Privileged Information for Efficient Image Super-Resolution},
booktitle={Proceedings of European Conference on Computer Vision},
year={2020},
}
Some parts of this code (e.g., data_loader) are based on EDSR-PyTorch repository.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent