Jiebin Yan¹, Ziwen Tan¹, Yuming Fang¹, Jiale Rao¹, and Yifan Zuo².

¹ School of Computing and Artificial Intelligence, Jiangxi University of Finance and Economics

• February 26, 2025: The arXiv version of our paper is released: https://arxiv.org/abs/2502.19046

• February 4, 2025: Our paper is accepted by Pattern Recognition!

• July 31, 2024: We upload the Max360IQ source code.

You can download databases at JUFE、OIQA and CVIQ

• Extract the viewports of omnidirectional images by using the tool getImageViewport

• use_gru(True/False): it is recommended to set True when there is a temporal relationship in the viewport sequence and loading the weights trained on JUFE
• Modify the load_ckpt_path to load pre-trained weights
• Modify the test_img_path to prepare the image data, the directory structure of a testing image is as follows:

Test_image/
├── vs1/
│   ├── vp1.png
│   ├── vp2.png
│   ├──   ...
│   ├── vpK.png
├── vs2/
│   ├── vp1.png
│   ├── vp2.png
│   ├──   ...
│   ├── vpK.png
├── ...
└── vsM/
    ├── vp1.png
    ├── vp2.png
    ├──   ...
    └── vpK.png

• Run the following code for inference one image

python inference_one_image.py

• The pre-trained weights can be downloaded at the Google drive
• Edit the config.py for an implement
• Run the file train.py and test.py for training and testing
• If you need train our model on other databases, loading weights pre-trained on JUFE could has better training results

@article{yan2024max360iq,
title={Max360IQ: Blind omnidirectional image quality assessment with multi-axis attention},
author={Yan, Jiebin and Tan, Ziwen and Fang, Yuming and Rao, jiale and Zuo, Yifan},
volume={162},
pages={111429},
year={2025},
}