search

zhmiao / OpenLongTailRecognition-OLTR

Pytorch implementation for "Large-Scale Long-Tailed Recognition in an Open World" (CVPR 2019 ORAL)
BSD 3-Clause "New" or "Revised" License
839 stars 128 forks source link

Can you explain the implementation of the disccentroidloss? #47

Closed tonysy closed 4 years ago

tonysy commented 4 years ago

Hi, thanks for your work. I am confusing about the implementation of the disccentroidloss, could you share the correct mathematical formulation of this loss function. I think the implementation is different from the paper details (equation. 9)

Thanks.

zhmiao commented 4 years ago

Hello @tonysy , sorry for the late reply! In Eq(9), basically it is positive the distance to corresponding centroids, and negative the distance to other centroids. The only difference is that we not only calculate the distance to centroids but also calculate dot product to centroids for repelling loss. Because we did normalization to the final feature, dot product is just the cosine similarity, another distance metric. We did not specify this in the paper. We will update the paper accordingly. Thanks again.

zhmiao commented 4 years ago

It has been 26 days since last comment. I will close this issue now. If you have any more questions, please reopen this issue. Thanks.