search

TreB1eN / InsightFace_Pytorch

Pytorch0.4.1 codes for InsightFace
MIT License
1.74k stars 423 forks source link

some questions to talk on evaluate and face_verify #62

Closed ruiming46zrm closed 5 years ago

ruiming46zrm commented 5 years ago

  1. why use Euclidean distance but not angular to calculate accurancy when evaluating lfw etc.

  2. why use mean values of embeddings of one ID as class_centre when registering, with not l2_norm to put on the hypersphere

nhatsmrt commented 5 years ago

For 1, do you mean cosine similarity? I noticed that he normalized the embedding to 1. This means that the cosine similarity is equivalent to (squared) Euclidean distance, plus a constant.

ruiming46zrm commented 5 years ago

For 1, do you mean cosine similarity? I noticed that he normalized the embedding to 1. This means that the cosine similarity is equivalent to (squared) Euclidean distance, plus a constant.

thank you very much, i get it