Not an issue per se: Why is CaffeMobile::SetMean used? How to manipulate it to get better result if at all possible?

[rivaldo4t]

Kindly let me know what this function does.

[sh1r0]

As name tells, set mean values for mean subtraction. And also, code tells you the truth. Thanks.

[rivaldo4t]

A little weak on the concepts here. Why is mean subtraction required in evaluation of a net?

[sh1r0]

IMHO, this is not weird for following the same preprocessing protocol as in the training.

[sh1r0]

Note: the "mean" here is not from your test image, it's the one used in the model training.

[rivaldo4t]

Yes I figured that. The mean values are set to some seemingly random values in the demo code.

float[] meanValues = {104, 117, 123}; caffeMobile.setMean(meanValues);

This is what does not make sense to me. What are these values.

[sh1r0]

https://github.com/BVLC/caffe/blob/master/models/bvlc_reference_caffenet/train_val.prototxt

[rivaldo4t]

Ok; not random. Still not clear on the WHY part.

[rivaldo4t]

Is it because we have trained the network on transformed images (mean subtracted) by these mean values and we must transform the test image in the same way during the evaluation phase?

[sh1r0]

Exactly.

[rivaldo4t]

Okay, thanks.