The performance of model in dataset glint360k

[Phil-Lin]

Hi ,Thank you for your work sharing. I added Adaface to project insightFace and verified the effect of Adaface in dataset MV1Mv3, but in Glint360K, the effect of Adaface is worse than that of Cosface. Have you verified it on dataset Glint360K?

dataset | method | backnone | LFW | CFPFP | AGEDB | IJBC(1E-4) MS1MV3 | Adaface | r100 | 99.867 | 99.014 | 98.3 | 97.17 MS1MV3 | Arcface | r100 | 99.85 | 98.9 | 98.55 | 96.85 Glint360k | Adaface | r100 | 99.83 | 99.15 | 98.45 | 97.38 Glint360k | Cosface | r100 | 99.817 | 99.2 | 98.65 | 97.55

[mk-minchul]

I have not run AdaFace in Glint360k. One potential improvement comes from AdaFace capable of handling strong augmentations unlike other margin based works (augmentations which may cause unidentifiable images and helpful for low quality dataset) as noted in Table 2 of the main paper. So when using the InsightFace, incorporating the training data augmentation in this repo into your training code may help.

[Phil-Lin]

Thanks for your reply. I have added the training data augmentation in this repo into insightFace project. And obtain the above experimental results.

[ducanhnguyen-lab]

@Phil-Lin Great work! Can you share your code which use AdaFace and InsightFace? Thank you <3