search

tztztztztz / eql.detectron2

The official implementation of Equalization Loss for Long-Tailed Object Recognition (CVPR 2020) based on Detectron2. https://arxiv.org/abs/2003.05176
Apache License 2.0
202 stars 17 forks source link

Implementation Details #4

Closed valencebond closed 4 years ago

valencebond commented 4 years ago

Thanks for your good paper and code. As in Implementation Details Section 4.2,

"we make a small modification when EQL is applied on LVIS ...... the weight term of Equation 7 will be 1 for those categories, even if they are rare ones."

In original EQL, the loss weight is 0 only for negative samples of foreground rare category . But according to Sec4.2, the weight term for negative samples of foreground rare category will be 1, That means all samples weight is 1?

i am not familiar with LVIS datasets and the settings of additional image-level annotations, so please correct me if there is something wrong.

tztztztztz commented 4 years ago

In original EQL, the loss weight is 0 only for negative samples of foreground rare category . But according to Sec4.2, the weight term for negative samples of foreground rare category will be 1, That means all samples weight is 1?

Yes, your understanding is right. BTW, this implementation doesn't apply this rule since we found that it doesn't lead to a better result.