I am trying to implement Knowledge Review for Object Detection using Faster RCNN as a Detector. I plan to use Resnet 50 as the backbone for the Teacher and Resnet 18 as the backbone for the student.
Both the Teacher and Student backbones will be having four layers in the backbone and similar architecture.
For Knowledge Review to work the shapes of the feature maps of the Student after the ABF should match the shape of the feature map of the Teacher.
What would be the strategy to achieve this? Do we need to set specific sizes for upsampling/ downsampling?
Hi,
I am trying to implement Knowledge Review for Object Detection using Faster RCNN as a Detector. I plan to use Resnet 50 as the backbone for the Teacher and Resnet 18 as the backbone for the student.
Both the Teacher and Student backbones will be having four layers in the backbone and similar architecture.
For Knowledge Review to work the shapes of the feature maps of the Student after the ABF should match the shape of the feature map of the Teacher.
What would be the strategy to achieve this? Do we need to set specific sizes for upsampling/ downsampling?
Thanks and Regards