search

hustvl / BMaskR-CNN

[ECCV 2020] Boundary-preserving Mask R-CNN
https://arxiv.org/abs/2007.08921
Apache License 2.0
189 stars 41 forks source link

What is different between Cascade BMask R-CNN and X101-FPN in detectron2 ? #11

Closed charlescho64 closed 3 years ago

charlescho64 commented 3 years ago

Hello,

I have question about your work. It seems X101-FPN (detectron2's result table) is better than R-101. If your model use X101-FPN baseline, is it much better? So if you didn't, is there some reason?

wondervictor commented 3 years ago

@charlescho64 Sure, ResNeXt-101-FPN performs better than ResNet-101-FPN, and the conclusion also holds in BMask R-CNN. Due to the computation and memory budget as well as the limitation of GPUs, we didn't adopt the ResNeXt-101-FPN.

charlescho64 commented 3 years ago

@wondervictor Thank you for your reply. your job is very impressive.

charlescho64 commented 3 years ago

@wondervictor Could you give me the way to use "RestNeXt-101-FPN" for training in your model(BMASK)?

wondervictor commented 3 years ago

@charlescho64 You can refer to this config file mask_rcnn_X_101_32x8d_FPN_3x.yaml which includes ResNeXt-101-FPN.

charlescho64 commented 3 years ago

@wondervictor Thank you.