Recognizing Families In the Wild Data Challenge (5th Edition) in conjunction with FG 2021 Forum

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

> train-pairs.csv file contains only positive pairs!

Hello,

The train-pairs.csv file contains only positive pairs! Do we have to randomly create negative samples randomly?

Kind regards,
Salah

Posted by: Bekhouche @ Dec. 3, 2019, 4:08 p.m.

Greetings Salah:

Thank you for pointing this out.

We only provide the positive member pairs, not a face list, as there are different ways to solve the problem. For instance, some may learn using a metric based on the separation of positive and negative, while others are based on classes (e.g., Sphereface-like, where one could train on the family IDs, as one solution).

With that, we did not want to misguide nor influence participants' mindsets too much-- we figured if we provided a list of samples that would push people to think of methods to fit the data, when the data is, in fact, the family data.

This is a bit of a rant, but I wanted to explain as this is a very good question and a potential source of confusion.

The validation list was provided for development purposes. Plus, participants are able to make successful submissions before the week of Phase III.

The test set will be a list very much like the validation, but without any labels or hints of family ID. At this time, the train + validation sets are available for training.

Thus, those not using negative samples can just remove them from the validation list when training for the final submission at the beginning of January. The same vice versa (those that need negatives, can generate them however they like with both training and validation). For instance, perhaps someone has a better way to generate negative samples that boosts their performance. We hope to see different approaches and different views.

Nonetheless, we almost released a sample train list with negative pairs (as part of a PCA demo or something). Would this be of interest to you? I can spin up a list with a fair distribution of member faces used as negative pairs if this will help your progress.

Thanks again for pointing this out!

~Joe

Posted by: jvision @ Dec. 3, 2019, 4:42 p.m.

Yes, I am interested in this sample list.

Posted by: Bekhouche @ Dec. 5, 2019, 8:10 a.m.

Okay. It is on my list of things to do today.

As soon as I add it to the competition I will notify via this forum.

Posted by: jvision @ Dec. 5, 2019, 12:46 p.m.

The sample face lists are available under Get Data:

Sample Train Face List only positives, https://1drv.ms/u/s!AkDk_XdBkJ9wgagGQ-OtQsibmD0gBw?e=hizTsn
Sample Train Face List with negative, https://1drv.ms/u/s!AkDk_XdBkJ9wgagCPB-UakHehdEACw?e=nfvDcu

Posted by: jvision @ Dec. 9, 2019, 7:45 p.m.

Thank you :)

Posted by: Bekhouche @ Dec. 10, 2019, 5:38 a.m.
Post in this thread