A modified GAP paradigm to improve class wise localization in images using convolutional layers only.
Original GAP - Follow http://cnnlocalization.csail.mit.edu/Zhou_Learning_Deep_Features_CVPR_2016_paper.pdf
Modified GAP - Using classification error change in absence of filter as weight for heatmap generation
Newer GAP - Using average excitation level of filter as weight for heatmap generation
Some illustrations of the working of Modified GAP are shown below
