if replace DCN with normal conv, How much performance drop does it make?

[zyc4me]

if replace DCN with normal conv, How much performance drop does it make? such as resnet18, AP 28.1 will drop to ??

[xingyizhou]

I haven't tried res 18. DLAv0 (The original DLA, without dcn and final upsample) is about 33 flip AP. You are welcomed to try the vanilla resnet yourself and report the accuracy.

[Wondersui]

if you change DCNv2 to DCNv1, the AP will drop to?

[xingyizhou]

@Wondersui I do have tried using DCN implementation from mmdet. Here DCNv1 and DCNv2 give the same performance. However, both implementations are slightly (< 1AP) worse than CharlesShang's DCN implementation. That's why we keep using this implementation. I am not sure if the difference is from randomness or details in the DCN implementation. To answer your question: the AP will drop at most 1.

[cwjhx]

I haven't tried res 18. DLAv0 (The original DLA, without dcn and final upsample) is about 33 flip AP. You are welcomed to try the vanilla resnet yourself and report the accuracy.

I will try DLAv0 （The original DLA, without dcn and final upsample），How train this model？Thank you.

[xingyizhou]

--arch dlav0_34

[wangshankun]

I haven't tried res 18. DLAv0 (The original DLA, without dcn and final upsample) is about 33 flip AP. You are welcomed to try the vanilla resnet yourself and report the accuracy.

I will try DLAv0 （The original DLA, without dcn and final upsample），How train this model？Thank you.

have any result？

[andriiD]

resnet18 with DCN 0.285. vanilla resnet with the same training settings and ordinary conv layers 0.231

[dheerajpreddy]

Can you please upload the models trained without DCN? Thanks

[andriiD]

Can you please upload the models trained without DCN? Thanks

Unfortunately I cannot share the model weights