Is there some difference in the baseline between Yours and PromptDet

hanoonaR / object-centric-ovd

[NeurIPS 2022] Official repository of paper titled "Bridging the Gap between Object and Image-level Representations for Open-Vocabulary Detection".

Apache License 2.0

284 stars 19 forks source link

Is there some difference in the baseline between Yours and PromptDet #14

Closed xushilin1 closed 1 year ago

xushilin1 commented 1 year ago

The baseline's $AP_r$ of your and PromptDet are 12.2 and 7.4, respectively. But both of you seem to replace the original classification with softmax cross-entropy of Mask RCNN with similarity score with sigmoid cross entropy. Why is your baseline much higher PromptDet?

Could you please give more details about the difference between your baseline and PromptDet's baseline?

Thanks for your reply.

hanoonaR commented 1 year ago

Hi @xushilin1 ,

For LVIS experiments, we use Mask R-CNN with ResNet-50FPN backbone with federated loss and sigmoid cross-entropy. We notice that the federated loss helps in improving the rare AP. Hope this helps.

Thank you.