search

endernewton / tf-faster-rcnn

Tensorflow Faster RCNN for Object Detection
https://arxiv.org/pdf/1702.02138.pdf
MIT License
3.65k stars 1.58k forks source link

Position of ROI Pooling #393

Open Scofieldtangsiwei opened 5 years ago

Scofieldtangsiwei commented 5 years ago

@endernewton I am using resnet 101. But I don't quite understand why the Roi Pooling is in between the _image_to_head and _head_to_tail. From my understanding, the Roi Pooling should be right after the whole backbone net, and the result of Roi Pooling should go right into the classification net. Please tell me where I am wrong. Thanks a lot!

rnsandeep commented 5 years ago

Its the way faster-rcnn was initially designed, another version which focuses on pooling from the final layer is called R-FCN. You can search for such implementations. You can read R-FCN paper by kiaming he.

Scofieldtangsiwei commented 5 years ago

@rnsandeep Thanks for your reply! I checked the paper again and I found the ROI Pooling was supposed to be performed before conv5 layer.