Negative scores

[shi...@yellresearch.com]

Hi, 

I am new to openBr. (please bear with me if my questions are silly)

To get started, I've followed the instructions here to build OpenBR on my unbuntu machine.

Then I tried out the command as written here:

http://openbiometrics.org/doxygen/latest/group__cli.html

For comparison using FaceRecognition algorithm, scores like 0.6, 0.7 make sense to me.

But then, I followed the example command to train the given dataset (downloaded via ./downloadDatasets.sh) and made comparison with one photo in ../data/family.jpg against the trained algorithm, I noticed some negative values (-7.xxx, -6.xxx etc).

So I did another test with the following commands:

./br -algorithm 'Open+Cvt(Gray)+Cascade(FrontalFace)+ASEFEyes+Affine(128,128,0.33,0.45)+(Grid(10,10)+SIFTDescriptor(12)+ByRow)/(Blur(1.1)+Gamma(0.2)+DoG(1,2)+ContrastEq(0.1,10)+LBP(1,2)+RectRegions(8,8,6,6)+Hist(59))+PCA(0.95)+Normalize(L2)+Dup(12)+RndSubspace(0.05,1)+LDA(0.98)+Cat+PCA(0.95)+Normalize(L1)+Quantize:NegativeLogPlusOne(ByteL1)' -train ../data/ATT/img FaceRecognitionATT

./br -algorithm FaceRecognitionATT -compare ../data/ATT/img/ ../data/ATT/img/s1 results.csv

 Based the output in the results.csv, if it is a exact match, then the value is 0, else some other negative scores. I am confused.

 

What does these negative values indicate? Are these values valid?

Or did I do it wrong?

Shinee

Ot

I f

I though

-

I did

Did

So

[Josh Klontz]

Shinee,

You didn't do it wrong -- it's just that the algorithm that you trained is slightly different (simplified) from FaceRecognition. One of the simplifications was to remove score normalization, so the score range is different but still entirely valid. You should be able to add score normalization back by replacing "NegativeLogPlusOne(ByteL1)" with "MatchProbability(NegativeLogPlusOne(ByteL1))" and retraining.

-Josh

--
You received this message because you are subscribed to the Google Groups "Open Source Biometric Recognition" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openbr-dev+...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

[shi...@yellresearch.com]

Thank you, Josh!

I can see why the algorithm is a simplified version. After retraining, I made another comparison and noticed that in general, the match scores are generally lower than the one using FaceRecognition (some even with "nan" score). I guess, until I can create a better one, I will stick with FaceRecognition. :)

Thanks again.