Where is Retinaface's accuracy sourced from?

[kartikeyporwal]

Hi team,

Thank you for releasing this amazing project.

I was skimming through your paper YOLO5Face: Why Reinventing a Face Detector and found that it's claiming to have better accuracy than RetinaFace and lists the RetinaFace's accuracy RetinaFace [29] ResNet50 [41] 94.92 91.90 64.17 29.50 37.59 that is different/lesser from what is listed in their official papers and repo.

RetinaFace has two papers 1. RetinaFace: Single-stage Dense Face Localisation in the Wild 2. RetinaFace: Single-shot Multi-level Face Localisation in the Wild

both shows better accuracy than what is displayed in YOLOV5-Face paper.

Also, Insightface's Retinaface's README shows a different accuracy of WiderFace validation mAP: Easy 96.5, Medium 95.6, Hard 90.4. on RetinaFace-ResNet50 model.

RetinaFace seems a winner here when we compare the accuracy with yolov5-face and even with newly released yolov7-face or am I missing something here?

Could you please clarify? I appreciate you response in this.

Thanks

[derronqi]

test size 640

[kartikeyporwal]

Thanks for your response. They have mentioned different accuracy for RetinaFace-ResNet50.