Open Lyan-ing opened 2 years ago
and how to sample the positives in paper section 4.4? thank you
From section 4.4: "To preserve fine-grained information we sample local features from within each mask and optimize them using the DetCon objective." This includes positive and negative samples.
We didn't evaluate linear classification on ImageNet as we were interested in transfer learning.
Got it. thank u.
How about the linear evaluation on ImageNet? ^_^