Its an interesting work, but the comparison seems a little unfair

prannaykaul / lvc

Official repo for our CVPR 22 paper: "Label, Verify, Correct: A Simple Few-Shot Object Detection Method"

Apache License 2.0

88 stars 14 forks source link

Its an interesting work, but the comparison seems a little unfair #4

Closed liuweijie19980216 closed 2 years ago

liuweijie19980216 commented 2 years ago

I noticed that KNN is trained using the pre-trained VIT model, which means that the work also uses training samples of pre-trained VIT, which seems unfair.

prannaykaul commented 2 years ago

Hello, we use a ViT pretrained with the self-supervised DINO method. This uses ImageNet-1k but without labels. So we do not use any extra labelled data.

liuweijie19980216 commented 2 years ago

Thanks for your reply. Actually, I love the work very much.