search

YueLiao / CDN

Code for "Mining the Benefits of Two-stage and One-stage HOI Detection"
Apache License 2.0
89 stars 15 forks source link

Confusion about parameter conversion #18

Closed ASMIftekhar closed 2 years ago

ASMIftekhar commented 2 years ago

Hello, Thanks a lot for your work. As I can see in the mentioned line, you are renaming DETR decoder layer to stage2_decoder. However, in your model there is no stage2_decoder. Does that mean you are not initializing the decoders from DETR?

https://github.com/YueLiao/CDN/blob/71c570296ecc7194c7abf99f5207344aa9c8dabe/convert_parameters.py#L50

YueLiao commented 2 years ago

Thx for your interest in our work! Thanks for your reminder, It is a missing update when we format our code. However, we find some researchers have achieved higher performance with this 'wrong setting', e.g., the cdn-s has achieved about 32 mAP. Thus, it may be better without initialization for stage_2.

ASMIftekhar commented 2 years ago

So I assume in your reported result in the paper you do initialize the decoder with DETR weights?

YueLiao commented 2 years ago

So I assume in your reported result in the paper you do initialize the decoder with DETR weights?

Yes, you are right. You could have a look at #4. The results reported by other researchers are all higher than the paper reported.

ASMIftekhar commented 2 years ago

Sorry for asking again. Among two decoders I guess you initialize the interaction decoder with detr weights?

YueLiao commented 2 years ago

Sorry for asking again. Among two decoders I guess you initialize the interaction decoder with detr weights?

Yes, just in our orignial experiments.

ASMIftekhar commented 2 years ago

Thanks a lot. Closing the issue.