This repository contains the official implementation for ICCV 2023 paper:
Weiran Gou, Ziyao Yi, Yan Xiang, Shaoqing Li, Zibin Liu, Dehui Kong, Ke Xu. [arxiv]
SYENet is an efficient network that could handle multiple low-level vision (isp, lle and sr) tasks. SYENet utilises re-parameterization for fast inference and got the highest score in MAI 2022 Learned Smartphone Challenge.
If you find our work useful in your research, please cite:
@InProceedings{Gou_2023_ICCV,
author = {Gou, Weiran and Yi, Ziyao and Xiang, Yan and Li, Shaoqing and Liu, Zibin and Kong, Dehui and Xu, Ke},
title = {SYENet: A Simple Yet Effective Network for Multiple Low-Level Vision Tasks with Real-Time Performance on Mobile Device},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2023},
pages = {12182-12195}
}
- python 3.8
- pytorch == 1.12.1
- numpy == 1.23.3
- cv2 == 4.7.0
- PIL == 9.2.0
- tqdm == 4.64.1
- yaml == 6.0
Edit the yaml files (isp.yaml, lle.yaml, sr.yaml) in ./config.
You are recommended to use basicsr to train our sr models to get higher PSNR. We put the train/test configuration files for training/testing our sr models using basicsr in ./config, which are sr_basicsr_train.yaml and sr_basicsr_test.yaml.
If you want to re-parameterize the model and save it, please set 'train
For isp and lle tasks, we utilise a warmup phase which is a self-supervised training stage. This phase could be cancalled by setting 'train
python main.py -task train -model_type original -model_task isp/lle/sr -device cuda
Set the 'model
If you are loading an original pretrained model, but you want to re-parameterize it before inference, you could set 'model
You could save the images generated in the test by setting 'test
python main.py -task test -model_type original -model_task isp/lle/sr -device cuda
You could save the images generated in the demonstration by setting 'demo
python main.py -task demo -model_type original -model_task isp/lle/sr -device cuda