ICCV 2023: Improving Lens Flare Removal with General Purpose Pipeline and Multiple Light Sources Recovery
Yuyan Zhou, Dong Liang, Songcan Chen, Sheng-Jun Huang, Shuo Yang
MIIT Key Laboratory of Pattern Analysis and Machine Intelligence, College of Computer Science and Technology,
Nanjing University of Aeronautics and Astronautics, Nanjing, China
Imaging Technology Group, DJI Innovations Co. Ltd., Shanghai, China
School of Computer Science, Nankai University, Tianjin, China
Wu el al. training images are provided in How to Train Neural Networks for Flare Removal (Wu et al. ICCV 2021)
Flare7K training images are provided in Flare7K: A Phenomenological Nighttime Flare Removal Dataset (Dai et al. NeurIPS 2022).
Please follow their instructions to access the data.
The consumer electronics dataset can be downloaded in Google Cloud or Baidu Netdisk (password:ptb3).
The inference code based on Uformer can be downloaded in Google Cloud or Baidu Netdisk (password: vmxl).
The backbone of our code is from Wu et al. How to Train Neural Networks for Flare Removal. We imporve the synthesis pipeline and light source recovery method to obtain better performance on flare removal.
python train.py --flarec_dir=path/to/captured/flare --flares_dir=path/to/simulated/flare --scene_dir=path/to/scene/image
python remove_flare.py --input_dir=path/to/test/image/dir --out_dir=path/to/output/dir --model=Uformer --batch_size=2 --ckpt=path/to/pretrained/model