search

WXinlong / DenseCL

Dense Contrastive Learning (DenseCL) for self-supervised representation learning, CVPR 2021 Oral.
GNU General Public License v3.0
544 stars 70 forks source link

2.3 negative sample #10

Closed wjczf123 closed 3 years ago

wjczf123 commented 3 years ago

I have 1 question and hope to hear from you: In section, 2.3 ''Each negative key t_ is the pooled feature vector of a view from a different image.'' Why not use the other parts of the two views of the same image as negative samples? This seems more make sense.

WXinlong commented 3 years ago

Please refer to Section 3.3 of the paper for details, i.e., the paragraph

Negative samples.

. https://arxiv.org/pdf/2011.09157.pdf

wjczf123 commented 3 years ago

Thank you for your reply. I saw this part. Have you tried using the other parts of the two views of the same image as negative samples? This seems more 'hard negative' than sample from a different image.

WXinlong commented 3 years ago

We didn't try that.

wjczf123 commented 3 years ago

OK. Thank you very much.

Holmes-GU commented 2 years ago

I have 1 question and hope to hear from you: In section, 2.3 ''Each negative key t_ is the pooled feature vector of a view from a different image.'' Why not use the other parts of the two views of the same image as negative samples? This seems more make sense.

So the 'pooled feature vector of a view from a different image' means the global average pooled vector of a view from a different image? If so, t_ is equal to the global view of a different image?