Question about verification protocol

[yjhong89]

Hi. I am YJHong and thanks for sharing your great work!

My questions are:

• Why using l2 metric instead of cosine similarity when doing the verification ?
• When I use cosine similairty for verification using your pretrained model (IRSE50 / mobilefacenet both), its performance even worse than VGGFaceNet (trained by simple soft-max). Is this due to cosine similarity ?
• About preprocess: 1) bbox using face detector - 2) plus margin - 3) resize to 112,112 - -4) ToTensor (divide by 255) - 5) Normalize (0.5,0.5,0.5) -> -> this right ?