What is the requirement of opinionScores for fitbrisque?

In the example:

https://www.mathworks.com/help/images/ref/fitbrisque.html#d126e112118

the opinionScores is set to random numbers in the range of [0 100], no problem. However, when modifing this to any identical values in the range [0 100], such as

opinionScores = 50*ones(1,size(imds.Files,1))

it showed error:

model = fitbrisque(imds,opinionScores')
Error using brisqueModel
Expected Alpha to be nonempty.
Error in brisqueModel>validateAlpha (line 298)
validateattributes(Alpha,images.internal.iptnumerictypes,{'nonempty','real', 'column' ...
    Error in brisqueModel (line 135)
    validateAlpha(Alpha);
    Error in brisqueModel.computeBRISQUEModel (line 222)
    obj = brisqueModel(svrmdl.Alpha, svrmdl.Bias, svrmdl.SupportVectors, svrmdl.KernelParameters.Scale);
    Error in fitbrisque (line 87)
    model = brisqueModel.computeBRISQUEModel(options.IMDS, options.SCORES);

How to implement the opinionScores values other than random?

Thanks.

Answer 1

[Walter Roberson]

It looks to me as if that should only happen if imds is empty.

[John]

imds is not empty. As soon as the opinionScores changes back to

opinionScores = 100*rand(1,size(imds.Files,1))

everything works fine.

[Walter Roberson]

It would help to have the content of imds to test with.

[John]

MatLab example, link in the question.

[Walter Roberson]

classreg.learning.svmutils.cpu.solve is a built-in function, and is returning alpha all 0 for the case of opinionScores being all the same.

A step in SVMImpl eliminates the alpha that are 0. With all the alpha having been returned as 0, that results in all the alpha being removed.

After that, brisqueModel complains that the alpha is empty.

Unfortunately, we cannot chase classreg.learning.svmutils.cpu.solve further.

[John]

In this case, how to set the values other than random?

[Walter Roberson]

If even one of the opinionScores is set differently by at least 0.25 then it works.