search

Sense-X / Co-DETR

[ICCV 2023] DETRs with Collaborative Hybrid Assignments Training
MIT License
968 stars 105 forks source link

encoder loss #78

Closed yimeng436 closed 10 months ago

yimeng436 commented 10 months ago

I don't understand why the label label is set to zero when the result of encoder is used to calculate the loss. Can you give me some guidance? Thank you very much

image

TempleX98 commented 10 months ago

I guess deformable-detr just follows the design of Faster-RCNN, which uses an RPN to classify the generated proposals into foreground or background.

yimeng436 commented 10 months ago

That is to say, here the all-zero label, zero represents the foreground ?

TempleX98 commented 10 months ago

That is to say, here the all-zero label, zero represents the foreground ?

Yes