DGQA

This is the official pytorch implementation of CVPR2024 "Bridging the Synthetic-to-Authentic Gap: Distortion-Guided Unsupervised Domain Adaptation for Blind Image Quality Assessment".

Requirement

• Python 3.6
• pytorch 1.4.0
• torchvision 0.5.0

Usage

1. Download the pretrained model and put them into model/.

   Google Drive: https://drive.google.com/file/d/12BZ0Xts5xTj0ppqyZRA4wFuIt6s3ONG2/view?usp=drive_link

   百度网盘：链接：https://pan.baidu.com/s/1Z9KjI6wXj6Kr1vy4YgyWew 提取码：9do9

2. To selecting the appropriate distortions for the target dataset, you need to run:

   python sel_dist.py --dataset target_dataset --root target_dataset_root
   

Results

The distortion types selected by DGQA on several common datasets are shown in the following table. Selecting only images with these distortion types in KADID-10K for training can significantly improve generalization performance and stability.

	#1	#2	#3	#4	#7	#9	#10	#11	#12	#16	#17	#18	#19	#21	#22	#23	#25
LIVEC	√		√				√				√	√					√
Koniq-10k	√		√							√	√	√					√
BID	√		√			√								√		√	√
PIPAL-1	√	√	√			√					√	√					√
PIPAL-2	√	√	√			√					√					√
PIPAL-3	√		√							√	√	√	√				√
PIPAL-4	√	√	√		√	√	√				√	√					√
PIPAL-5	√	√	√	√		√
PIPAL-6	√	√	√			√	√	√	√	√	√	√	√	√	√		√

For LIVEC and Koniq-10k, these distortion types are more recommanded, which are selected based on a greedy selection strategy.

	#1	#2	#3	#9	#10	#17	#18	#20	#25
LIVEC	√		√	√			√	√	√
Koniq-10k	√	√			√	√		√	√

Citation

If you find our code helpful for your research, please consider citing:

@inproceedings{li2024bridging,
  title={Bridging the Synthetic-to-Authentic Gap: Distortion-Guided Unsupervised Domain Adaptation for Blind Image Quality Assessment},
  author={Li, Aobo and Wu, Jinjian and Liu, Yongxu and Li, Leida},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={28422--28431},
  year={2024}
}