ECCV 2020 ChaLearn Looking at People Fair Face Recognition challenge Forum

Go back to competition Back to thread list Post in this thread

> Use of additional training data or pretrained representations

I am planning to work on the problem and participate in the challenge. However, there are two questions on my mind:
(1) use of any additional training data,
(2) use of pretrained representations

I could not find any statement on the website. In my opinion, the challenge must be solely on the provided dataset and w/o using any pretrained network in order to understand the contribution of proposed methods to the fairness measures. Could you clarify those points, please?

Posted by: omer.sumer @ April 12, 2020, 11:51 a.m.

Dear omer.sumer,
we agree that would be interesting to verify the effectiveness of submitted methods by limiting participants to only use the provided train data with no pretrained models. However, due to the format of the challenge (i.e., submission of results, and no code submission), it would be difficult to verify both constraints, even in the code verification stage. For this reason, among others, we did not impose any constraint with respect to external data or pretrained models. We believe this decision still allow participants to provide interesting solutions for the problem. Please, don't hesitate to contact us if you have any doubt/problem. Best regards, the organizers.

Posted by: juliojj @ April 12, 2020, 3:49 p.m.

Hi,
1. Is it allowed to use the original IJB-C dataset as training set?
2. If I used some privated algroithms(face detection and landmark deection) that are not allowed to open source , is it allowed to just provide the preprocessed test data at code verification stage?
3. If I used privated training framework, how should i do to submit training code?
4. If I used some privated training set, there is any problem at code verification stage?
Thank you.

Posted by: mcga @ June 29, 2020, 5:57 a.m.

Dear mcga,
A) All provided code and data should be accessible to organizers and for the public, with research sharing purpose permissions in any case (no need of commercial or extended permissions).
B) if the case A is not satisfied still you can provide all necessary instructions and access to private software and data (or pretrained models with clear specifications of used data and evidences) to the organizers to replicate your results and then to appear in the leaderboard and in challenge dissemination. Your info will appear but you will not be granted with prizes if you are a top entry;
C) if neither A nor B can be justified, your entry will be removed from any dissemination activity.
Best,
The organizers

Posted by: juliojj @ June 29, 2020, 9:45 a.m.
Post in this thread