what is the training dataset of provided ckpt?

[order-a-lemonade]

hi, thanks for you interesting work!!
i test some custom data, and got about scroe=35, but i am not sure this is high or low.
as i know, the output score is corresponding to the training dataset labels, right?

[Yliu724]

Hi! Thanks for your interest in our work and for testing with your custom data.

To address your question:

1. Our algorithm evaluates the difference between the input image and a high-quality reference image set. A higher score generally indicates a greater difference, meaning the image quality is lower. Image quality is usually relative, so to give you some context: in our experiments with the LIVE database, the quality score for the reference image "bikes.bmp" is 19.2337, while the score for a distorted image with Gaussian blur, "img10.bmp", is 42.7782. Based on this, if your test image has a quality score of 35, visually it should be of better quality than "img10.bmp" but not as good as "bikes.bmp".

2. It's also important to note that our algorithm is opinion-unaware. This means that during training, we only use a dataset of high-quality images without any opinion scores or labels to guide the model's learning.

I hope this helps clarify things! If you have more questions or need further clarification, feel free to ask.