Indeed, comparing the results of your algorithm to the results of other researchers is difficult. One way to achieve comparability without needing to re-implement the other algorithms, is to stick to well-defined evaluation protocols. For the ARface database, we have implemented a Python interface for a default evaluation protocol, which splits up the data into three sets: A training set (to estimate the data distribution), a development set (to optimize parameters of your algorithm) and an evaluation set to finally report the unbiased results of your algorithm. This protocol interface can be downloaded here:

https://pypi.python.org/pypi/bob.db.arface

At the same time we have implemented a library to run comparable face recognition algorithms on several databases, which will always follow the default evaluation protocols. A set of FR algorithms are already implemented, and adding new algorithms should be straightforward:

https://pypi.python.org/pypi/facereclib

@Steve: Please do not recommend to use the AT&T (aka. ORL) database to run experiments on them any more. Already in 2005, Anil K. Jain mentioned that the results on this database have saturated (actually in exactly the book than you've mentioned).

Best regards

Manuel

Hello,

If you want to compare the two algorithm, why not realize their experiment on your database of 50 people ?

Thanks David But I think Coding of all of that methods needs so much of time.

Hi Mehdi,

Researchers usually like to be cited. So, can you ask the authors to provide you the pictures they have used so that you apply your algorithm on the same database?

Good luck

Hello dear Mohammad I have the database. I don't think they would provide me with their codes.

Hi, there are some international databases to prove the efficiency and efficacy of these kind of algorithms.

Dear Mehdi

Three databases to try are:

1) P. J. Phillips, H. Wechsler, J. Huang, and P. Rauss, “The

FERET database and evaluation procedure for face recognition

algorithms,” Image and Vision Computing, vol. 16, no. 5, pp.

295–306, 1998.

You can download this at http://www.itl.nist.gov/iad/humanid/feret/feret_master.html.

2) A.Martinez and R. Benavente, “The AR Face Database,” Tech.

Rep., Univerzitat Autonòma de Barcelona, 1998.

See http://www2.ece.ohio-state.edu/~aleix/ARdatabase.html.

3) The ATT database, cited in Anil K Jain and Stan Z Li, Handbook of face recognition,

Springer, 2005.

See http://www.cl.cam.ac.uk/research/dtg/attarchive/facedatabase.html.

Hope that this helps,

Cheers

Steve

Dear Mehdi,

To be sure for the  comparison

you can use and apply the algorithm for the same database.

www.airccse.org/journal/sipij/papers/3212sipij02.pdf

Regards

Hi. Dear Mehdi,

You can use databases FERET.

http://www.itl.nist.gov/iad/humanid/feret/feret_master.html

Look the link. May be useful.

Best regards, Shafagat

Indeed, comparing the results of your algorithm to the results of other researchers is difficult. One way to achieve comparability without needing to re-implement the other algorithms, is to stick to well-defined evaluation protocols. For the ARface database, we have implemented a Python interface for a default evaluation protocol, which splits up the data into three sets: A training set (to estimate the data distribution), a development set (to optimize parameters of your algorithm) and an evaluation set to finally report the unbiased results of your algorithm. This protocol interface can be downloaded here:

https://pypi.python.org/pypi/bob.db.arface

At the same time we have implemented a library to run comparable face recognition algorithms on several databases, which will always follow the default evaluation protocols. A set of FR algorithms are already implemented, and adding new algorithms should be straightforward:

https://pypi.python.org/pypi/facereclib

@Steve: Please do not recommend to use the AT&T (aka. ORL) database to run experiments on them any more. Already in 2005, Anil K. Jain mentioned that the results on this database have saturated (actually in exactly the book than you've mentioned).

Best regards

Manuel

Thanks Manuel you really understood the problem here and you gave the best answer.

Thanks again

Unfortunately, there are still many authors of face recognition papers (I guess that one that you mentioned in your original post is no exception) that do not understand the problem of comparability. Usually, they develop their own protocols, i.e., they split the data in a non-reproducible, most often random, way; or even worse: they perform several random splits and report results on the splits that fits their needs best.

Nowadays, good image databases (like the Labeled Faces in the Wild) come with pre-defined evaluation protocols so that results on these databases are generally comparable. However, the authors of the AR face database do not impose any protocol, so we have come up with our own. But, in opposition to most other researchers, we share our protocol. :-)

For comparison, obviously you need code or need to implement it. Most of the researchers will get the code in INTERNET or they will implement it for comparison. If it is not possible in your case, why don't you test it on selected images by you or on entire AR database or other face databases such as FERET, CAS-PEAL, LFW, CMU-PIE, FRGC, YALE database and etc. Show the obtained face recognition rates by your algorithm on these databases and explain  how better your algorithm performs on variety of face images such as frontal, different lighting conditions and pose changed faces. If possible try to mention how better your algorithm perform better than existing methods in terms of speed, computational complexity and accuracy. Give few conclusions by all your useful efforts which you have put in this project, also mention how you  are going to improve it in future.

Wishing you a good luck.....