Hello guys,
a few questions here.
In order to estimate age on the faces in images, our software
1)finds a face using a face detector
2) aligns the face using face landmark detector
3)performs the age recognition
Does it mean that in order to participate in the challenge, we must supply all the source code for the face detector, face landmark detector and age recognition? What if one used an opencv detector, should he submit the sources code of opencv?
It would be great if you provided landmark positions (at least eyes, mouth) for the faces, such that one can participate even without an opensource detector and alignment software.
Regards
Radek
Posted by: RadekSvoboda @ June 17, 2015, 11:27 a.m.I totally agree with this concern.
At least it is needed the center of the eyes to start the PRE-PROCESSING part.
A bad face localization affects directly the outcome of the system and you are not scoring the whole pipeline, right?
Posted by: basuam @ June 17, 2015, 4:26 p.m.Dear participants, thank you for your questions.
Regarding the code it is not a problem if you use OpenCV or other implementations at any stage of your system. At the end, we need your code to verify that last results on the test set are correct. In order to win the prizes your final code has to be opensource as defined in the rules. However, you can decide to not do that, you will figure as a winner but you will not receive the prizes. In any case we need your code for final verification and we have to be able to run it.
Regarding the different steps of your method, we totally agree that for most age recognition systems alignment and pre-processing are required. We explicitly did not include those annotations since we consider that those steps are part of the challenge.
If you have any more doubts please let us know.
Have a fun competition and good luck
PS: sorry for the problems with the provided data, we are fixing it.
A question on the code release. What if the end-to-end system is a pipeline concatenation of Matlab, Python, R and IPython Notebooks in both Windows and Linux? In this case, running the whole pipeline may be a convoluted process. Shall we assume that you do not worry about how hard is to run the pipeline?
Posted by: cristian.canton @ June 18, 2015, 4:47 p.m.Dear participant,
Do not worry about that but we would appreciate a summary about how to run such a complex system.
Hello,
A question here. Imagine you have a concatenation of systems like: 1) face detection, 2) face alignment, 3) Age estimation. For 1) and 2) you can provide the paper citation where the algorithm came from but you are not allowed to share the source code for whatever reason. At this point would it be acceptable to provide the output of modules 1) and 2) so that you can reproduce the results without having the original source code? Otherwise this competition involves too many sub-problems that are, per se, worth of a competition. Otherwise, we will be constrained to use 1) and 2) from open implementations (not a big problem either).
Thanks!
Posted by: cristian.canton @ June 30, 2015, 7:44 p.m.Dear,
In that case you can use any available code, but in addition to the outputs of the first stages we also need a link so that we can run first stages code to verify them. Then you can provide your own code for the last stage, only sharing that code, it will be not a problem.
Hope it helps, any doubts please ask
The last answer is in completely different spirit to the one I received. The organizers should really clarify what is and what is not a part of this competition, which modules require source code and which can be provided in (potentially encrypted) binaries and how this influences the final rankings / prizes.
Our situation is basically the same, we can provide binaries for face detection and registration (encrypted), and an evaluation source code for age recognition with pre-trained binary model. However it is not clear to me how the organizers will be able to verify anything using this pipeline, if, as far as I understood the challenge, the test data will be released to the public. In extreme case the age binary model can store the whole image database with human labels and it wont be possible to verify this.
Posted by: RadekSvoboda @ July 1, 2015, 5:42 a.m.Dear
Sorry for any confusion. What I commented in previous answers is that the whole code has to be tested by organizers. What I mean is that the whole pipeline has to be implemented, but you CAN use any other public code for a particular stage of the pipeline, referring to that code so that we can use the external code and replicate your outputs at any stage of the pipeline. If there exists no way we can verify each step it will be not considered for prizes.
We are now going to crop images so that you only perform age recognition, you have to implement the whole pipeline but you CAN use any existing data and code available (based on that the organizers will be able to verify the results). You can then submit the final opensource code and refer to any external code within your code (and it will be considered for prizes).
Of course there exists limitations, just one main file calling for external methods for ALL steps of the pipeline. This extreme case will not be considered for prizes (but still I am pretty sure that with that technique it would be very difficult to win the challenge).
On the other hand in the case that no opensource code is delivered at the end of the competition you will figure as winner / no winner with your rank value but you will be not eligible for prizes, as defined in the rules.
Sorry for the confusion, these are common issues about data and code, be free to comment any other doubt you may have.
PS: if you have further doubts regarding your code that can be of interest for all participants please post a message here. If you have further comments and doubts about your particular code you can mail me directly to: sergio@maia.ub.es
Again, have a fun competition
best