Awesome
KonIQ++: Boosting No-Reference Image Quality Assessment in the Wild by Jointly Predicting Image Quality and Defects
Source code for the BMVC'21 oral "KonIQ++: Boosting No-Reference Image Quality Assessment in the Wild by Jointly Predicting Image Quality and Defects".
The KonIQ++ database can be found here.
Our pretrained model can be found here.
Installation
Install dependencies
pip install -r requirements.txt
The code also requires apex, please refer to this repository for instructions.
Usages
Training & Testing on IQA databases
To reproduce reported results in the paper, first download the koniq++database.csv file from KonIQ++ Webpage and put it in your own path for KonIQ-10k dataset, then run:
python main.py --dataset KonIQ-10k --resize --lr 1e-4 -bs 8 -e 25 --ft_lr_ratio 0.1 --loss_type norm-in-norm --p 1 --q 2 --koniq_root 'YOUR OWN PATH'
Model will be trained and tested on the KonIQ++ dataset.
If you want to cross test on either CLIVE or SPAQ dataset, add following options:
--test_on_clive: Add option to cross test on CLIVE dataset after training on KonIQ++.--clive_root 'YOUR OWN PATH': Root path for CLIVE dataset.--test_on_spaq: Add option to cross test on SPAQ dataset after training on KonIQ++.--spaq_root 'YOUR OWN PATH': Root path for SPAQ dataset.
Note: We have conducted user screenings to the released version KonIQ++ labels, so the results might have small fluctuations regarding to the reported performances. To perfectly reproduce performances in the paper, you can use the Koniq++_old.csv label in the old_label folder as substitution.
Predicting Single Image Quality and Visualizing
To predict single image quality, run:
python test_image.py --root_path 'IMAGE PATH' --img_name 'TEST IMAGE NAME' --resize --p 1 --q 2
To visualize feature heatmaps from two side networks, add option --save_heatmap. Heatmaps will be saved under the image root path.
Citation
If you find this work useful for your research, please cite our paper.
Acknowledgements
This code is largely implemented based upon norm-in-norm IQA, we thank them for their sharing.