image-quality-assessment
image-quality-assessment copied to clipboard
Wrong inferred distribution using 'get_labels.py'
Using the settings in the current version of 'get_label.py' (i.e., 'samplespace' is [0:1:10] and algorithm='CG'), then if we input 'mean=7', the return distribution is wrong (i.e., with all the probabilities equal to 0.1). We can check this by printing Model.expections(), which will somehow be 4.5, far from what we expect (i.e., 7).
FIY: three ways to solve this problem:
- set 'samplespace' to [1:1:10];
- set 'algorithm' to 'BFGS';
- manually modify the mos that equals to '7' (e.g., changing the score to '7.1').
Hi,
Any developers checking on this issue? I mean if the same get_label.py code was used to generate the label for AVA image label, (since TID 2013 image label is a distribution) I suscept there could be a performance issue with the pre-trained model.