problem caused by dataset changing about cnniqa HOT 6 CLOSED

@salmora I'm not the author. I just implemented CNNIQA since the original author did not provide it.

for the 3D case, the inputs should be two images, you should change the network architectures so that it can accept your inputs and output the quality score.

You can refer the code in data_info_maker.m for data preparation, specifically, save the .mat file with -v7.3.

multiple (K） train-test splits are considered for avoiding content bias.
Larger K will be better and K has an upper bound, i.e., the number of different kinds of splits on the dataset.
In the paper, the author used K=100. And In my implementation, I generated K=1000 splits.
Each exp_id setting corresponds to an experiment under the exp_id%K-th split.

The code is very simple. Please read the whole repo and the paper, after that, I believe you can understand this method by yourself.

from cnniqa.

@salmora Please refer to LIVE dataset.

"Download realigned subjective quality data here."

After deleting the corresponding values for reference images (i.e., orgs=1), dmos_new and dmos_std are the subjective_scores and subjective_scores_std.

You should also read the paper that describes the dataset.

from cnniqa.

in my database for 3D images I don't have the information about subjective_scores_std I have only subjective_scores please I wanna know if that will effect the result of the modele

from cnniqa.

@salmora If you read the whole repo, you will find that in this repo subjective_scores_std is only used for calculating the performance criterion outlier_ratio, i.e., OR (line 75 in main.py).

Suggestion:
Make sure you have read the whole repo, the full paper about the model CNNIQA, and the full paper about the datasets.

from cnniqa.

Thank you very much for taking the time to answer my question, I have another one please how did you compute subjective_scores_std in your file LIVEfallinfo.mat I tried to extract this information from the database but I did not find enough information .
Thank you.

from cnniqa.

thank you so much

from cnniqa.