pseudo-object segmentations on output of convnet

gistfile1.lua

   local mask = function (inputImage, mask, th)
      local temp
      mask = mask[1]
      k = image.gaussian{size = 7, normalize = true}:float()
      mask = image.convolve(mask, k, 'same'):repeatTensor(3,1,1)
      temp = mask:gt(th):float():mul(.8):add(.2)

      return inputImage:clone():cmul(temp)
   end

   local gradManipulation = function (grad, th)
      return grad:float():abs():max(1):repeatTensor(3,1,1)
   end

   display.forward = function(output, fps)
      win:gbegin()
      win:showpage()

      -- compute heatmap for category person!
      local target = output[1]:clone():zero() -- only pick largest scale!
      target[2] = (output[1])[2]:clone():mul(-1) -- only pick largest scale!
      local gradInput = network.model:updateGradInput(source.scaled, target)

      heat = mask(source.frame, gradManipulation(image.scale(gradInput, source.frame:size(3), source.frame:size(2)), 0), .2)

      image.display{image = heat, win = win}

      -- end displaying!
      win:gend()
   end

This function creates a pseudo-proto-object segmentation from CNN output
output is output of CNN