About position-sensitive score maps

[qianyeqiang]

How the position-sensitive score maps generate? I don not find anything in detail in paper and author just say that use a bank of specialized convolutional layers as the FCN output.

Thank you very much.

[liyi14]

Hi, @supersuper-qian , I'm not sure whether I understood your question. The position-sensitive score map is the output blobs of typical convolutional layers (e.g. rfcn_cls, rfcn_bbox). While the position-sensitive ability is obtained by the assemble strategy in PSRoIPooling layer. You can review figure 1 and figure 3 in paper for more details.

[chengshuai]

Hi @liyi14 @YuwenXiong

In the paper,append a sibling 4k^2-d convolutional layer for bounding box regression,but in the code, 84^2-d for bounding box regression. What is the reason?

[liyi14]

@chengshuai where is the 84^2 you mentioned?