positive bias.py

[zhengchaobing]

Hi, I'm confused about this code in positive bias.py,

First, why did you do x/xn before geting w, not the same as the one introduced in the paper.

Second, ID is the result of a box filter applied to the input followed by downsampling, which is subsequently smoothed by an  approximate 2D Gaussian filter,  but you only use a pooling replace it.

[farazsaeedan]

dividing x by xn is just for numerical stability and does not change the value of the output result as it cancels out. for the DPP complete implementation look at DPP_asym_lite.lua in visinf _package.