This repository is the official MIPI Challenge 2023 implementation of Team LVGroup_HFUT in Nighttime Flare Removal.

The restoration results (Extraction Code：a9zs) of the tesing images and pretrained model (Extraction Code：l9zt) can be downloaded from Baidu Netdisk.

cd your/script/path

python infer.py --data_source your/dataset/path --model_path ../pretrained/epoch_0090.pth --save_image --experiment your-experiment-name

cd your/script/path

python train.py --data_source your/dataset/path --experiment your-experiment

The format of the dataset should meet the following code in datasets.py:

self.img_paths = sorted(glob.glob(data_source + '/train' + '/Flare' + '/*.*'))

self.gt_paths = sorted(glob.glob(data_source + '/train' + '/Flickr24K' + '/*.*'))

or

self.img_paths = sorted(glob.glob(data_source + '/val' + '/input' + '/*.*'))

data_source is given by the command line.

mode can be 'train' or 'val'.

when training and validating: the default path is '../results/your-experiment'

when testing: the default path is '../outputs/your-experiment/test'

when inferring: the default path is '../outputs/your-experiment/infer'