search

weiliu89 / caffe

Caffe: a fast open framework for deep learning.
http://caffe.berkeleyvision.org/
Other
4.77k stars 1.68k forks source link

Why is negative mining supported only if share_location=true? #473

Open yijunCai opened 7 years ago

yijunCai commented 7 years ago

Hi @weiliu89 I‘m wondering why is negative mining supported only if share_location=true, even when mining type is MAX_NEGATIVE. I've roughly read the code in MineHardExamples(). There might be some restriction for HARD_EXAMPLE, but it seems that MAX_NEGATIVE can work even if share_location=false. By the way, did you set share_location=true when training SSD for PASCAL VOC? Is that means all the 20 calsses in PASCAL VOC share the same location regressors?

weiliu89 commented 7 years ago

Only share_location=true is tested, which means all classes uses the same location regressor (not like Faster R-CNN).

yijunCai commented 7 years ago

I see. Thanks for your reply!