Regarding the rof dataset, the neutral/samuel_l._jackson/000051.jpg data that you have specified in your tests appears to be corrupted and does not seem correct for evaluation. Thank you in advance.
Posted by: kroobose @ March 10, 2024, 3:09 p.m.Many images in rof have some problem, such as no person (sunglasses/ben_mendelsohn_wearing_sunglasses 15-17, ...), wrong cropped (neutral/justin_theroux/000017.jpg, ...). About one hundred problem images in total.
Posted by: haohaohao @ March 11, 2024, 9:16 a.m.I'm facing the same issue where MTCNN does not detect any faces/landmarks so I can't apply my alignment algorithm to those images. Did you manage to find any workaround to that? I tried lowering the three MTCNN thresholds but then it finds non-existing faces.
Posted by: nicolo.didomenico @ March 11, 2024, 3:38 p.m.Many thanks for highlighting this point. We have carried out an in-depth analysis of the ROF comparisons. In total, the number of wrong comparisons is around 1%. Anyway, in order not to affect this point to the challenge, we have decided to send all participants a file with the corresponding wrong comparisons and the correct score you all should consider in your submission (i.e., '0' for non-mated comparisons and '1' for mated comparisons). This way we can avoid this problem. Many thanks and sorry about the inconvenience.
Posted by: FRCSyn @ March 13, 2024, 5:17 p.m.I think there are more wrong images in addition to the wrong comparisons sent in the email, and I have sent them to the official email.
Posted by: haohaohao @ March 18, 2024, 2:01 a.m.Many thanks for your collaboration in this regard. We have sent a new email adding the comparisons that you have indicated to us. Sorry for the inconvenience.
Posted by: FRCSyn @ March 19, 2024, 12:05 p.m.