search

unsky / FPN

Feature Pyramid Networks for Object Detection
524 stars 263 forks source link

as_rois.py & simple questions.. #22

Open hangil55 opened 6 years ago

hangil55 commented 6 years ago

First of all, Thank you for sharing FPN-caffe version I have some simple questions. 1) You apply "as_rois layer" in model/test.prototxt. However, It seems that you did not apply "as_rois_layer" in model/train.prototxt why this? 2) fc layers are weight sharing for this version. however, this is time consuming because when fc/h2 is done and then fc/h3 starts....(fc/h2 , fc/h3, fc/h4, fc/h5 each is done and merge for socre & bounding box) I think it is better gathering roi_pooling results (roi_pool/h2, h3, h4, h5) and forward fc layers. (i.e. merge ->fc // your version : fc2,3,4,5 -> merge) this is better when training & testing... Thank you!

unsky commented 6 years ago

as_rois layer assign rois to different level-p in test phase, proposal_target layer complete this task in train phase,

unsky commented 6 years ago

@hangil55 yes, the strategy your mention can work good, my work is loyalty to the original paper. and you can optimize it by yourself.

foralliance commented 6 years ago

@hangil55 @unsky i think hangil55 proposed schemeI gathering roi_pooling results (roi_pool/h2, h3, h4, h5) and forward fc layers. (i.e. merge ->fc // your version : fc2,3,4,5 -> merge) is megred rcnn framework That's my understanding, I'm not sure it is right.