Hi,
in the train-pairs.csv there are some inconsistencies with the brother/sister/siblings labels.
For example, lines 6/7 in the .csv file show that MID3 and MID4 are sisters and brothers at the same time.
F0001/MID3,F0001/MID4,ss
F0001/MID3,F0001/MID4,bb
When considering also the gender as given by F0001/mid.csv it turns out both are male.
This inconsistency is very frequent in train-pairs.csv (and val-all-face-pairs-no-label.csv) and also involves siblings.
I wrote a script to remove the wrong labels which reduces the number of train pairs from 264k to 205k (71k to 50k on validation).
Can you confirm my findings?
Regards,
Stefan
Let us check this today
Posted by: jvision @ Dec. 10, 2019, 2:33 p.m.Let us check this today
Posted by: jvision @ Dec. 10, 2019, 2:33 p.m.Thank you kindly for pointing this out, as you are indeed correct here. The bb, ss, sibs were all siblings, but the type of sibling was buggy (to say the least).
I can think of cases where one would only care about the type independent of gender (i.e., just sibling or not sibling). However, more cases come to mind where the gender would be useful to further separate types as ss, bb, sibs (mixed gender). Thus, updated tables were created and will be at the bottom of the Get Data section of this portal.
The train tables were updated with an added column ('old_tags'), which contains the values previously; now, 'ptype' column should be correct. I left the old so it is clear that only bugs were for sibling genders + just for the record. Ignore 'old_labels' columns unless needing to look at the old version of tags for whatever reason.
Thanks again!
Posted by: jvision @ Dec. 13, 2019, 10:25 p.m.same for val
Posted by: jvision @ Dec. 13, 2019, 10:53 p.m.