No penalty is seen in the final code when ID score is very low, which is unfair to the pursuit of high fidelity, may I ask what the penalty is?
Posted by: terrence @ March 20, 2025, 2:28 a.m.In the "Participant - Get Data - Identity Consistency" section, we specify:
"The submitted models should not have more than 10 pairs of images with mismatched IDs compared to the original images (tested on Test Data). Models that fail the AdaFace identity consistency test will not be included in the final score table. The goal of the threshold setting is to maintain basic identity consistency. Models that prioritize higher non-reference IQA metrics without considering identity consistency will be excluded at this step."