graph conv net ontop of Faster RCNN in mmdetection

[jamesbc123]

Could you suggest where I can implement a graph convolutional network? I would like to take the features from Faster R-CNN and pass them through a graph conv net. I'm unsure if I can do this in a new class (new ROI Head) or if in mm-lab there is some functions available elsewhere for this. I'm also unsure how the computational graph works in mmdetection: if we take the features from Faster R-CNN and pass it through another net, is this other net still able to be updated in back prop?

Thanks for your help

James

[jamesbc123]

Hello @hhaAndroid, do you have any thoughts on this?

Thank you!