STAR-pytorch

Implementation for paper “STAR: A Structure-aware Lightweight Transformer for Real-time Image Enhancement” (ICCV 2021).

CVF (pdf)

STAR-DCE

The pytorch implementation of low light enhancement with STAR on Adobe-MIT FiveK dataset. You can find it in STAR-DCE directory. Here we adopt the pipleline of Zero-DCE ( paper | code ), just replacing the CNN backbone with STAR. In Zero-DCE, for each image the network will regress a group of curves, which will then applied on the source image iteratively. You can find more details in the original repo Zero-DCE.

Requirements

  • numpy
  • einops
  • torch
  • torchvision
  • opencv

Datesets

We provide download links for Adobe-MIT FiveK datasets we used ( train | test ). Please note that we adopt the test set splited by DeepUPE for fair comparison.

Training DCE models

To train a original STAR-DCE model,

cd STAR-DCE
python train_dce.py 
  --lowlight_images_path "dir-to-your-training-set" \
  --parallel True \
  --snapshots_folder snapshots/STAR-ori \
  --lr 0.001 \
  --num_epochs 100 \
  --lr_type cos \
  --train_batch_size 32 \
  --model STAR-DCE-Ori \
  --snapshot_iter 10 \
  --num_workers 32 \

To train the baseline CNN-based DCE-Net (w\ or w\o Pooling),

cd STAR-DCE
python train_dce.py 
  --lowlight_images_path "dir-to-your-training-set" \
  --parallel True \
  --snapshots_folder snapshots/DCE \
  --lr 0.001 \
  --num_epochs 100 \
  --lr_type cos \
  --train_batch_size 32 \
  --model DCE-Net \
  --snapshot_iter 10 \
  --num_workers 32 \

or

cd STAR-DCE
python train_dce.py 
  --lowlight_images_path "dir-to-your-training-set" \
  --parallel True \
  --snapshots_folder snapshots/DCE-Pool \
  --lr 0.001 \
  --num_epochs 100 \
  --lr_type cos \
  --train_batch_size 32 \
  --model DCE-Net-Pool \
  --snapshot_iter 10 \
  --num_workers 32 \

Evaluation of trained models

To evaluated the STAR-DCE model you trained,

cd STAR-DCE
  python test_dce.py \
  --lowlight_images_path  "dir-to-your-test-set" \
  --parallel True \
  --snapshots_folder snapshots_test/STAR-DCE \
  --val_batch_size 1 \
  --pretrain_dir snapshots/STAR-ori/Epoch_best.pth \
  --model STAR-DCE-Ori \

To evaluated the DCE-Net model you trained,

cd STAR-DCE
  python test_dce.py \
  --lowlight_images_path  "dir-to-your-test-set" \
  --parallel True \
  --snapshots_folder snapshots_test/DCE \
  --val_batch_size 1 \
  --pretrain_dir snapshots/DCE/Epoch_best.pth \
  --model DCE-Net \

Citation

If this code helps your research, please cite our paper ?

@inproceedings{zhang2021star,
  title={STAR: A Structure-Aware Lightweight Transformer for Real-Time Image Enhancement},
  author={Zhang, Zhaoyang and Jiang, Yitong and Jiang, Jun and Wang, Xiaogang and Luo, Ping and Gu, Jinwei},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={4106--4115},
  year={2021}
}

GitHub

View Github