Triplet Training procedure?

[FrederikWarburg]

Hi,

Really interesting approach! I've worked a bit with the triplet loss recently, and I really think this approach seems great. However, it is not clear to me what hard-negative mining procedure you use for the triplet mining? In the paper you refer to ". End-to-end learning of deep visual representations for image retrieval." for the HNM procedure, but I can't find the description of the procedure here either. Could you elaborate a bit on the HNM?

Thank you in advance!