Is the loss based on "comparing 2 maps" or are you subtracting the exact coordinates ?
In my experience I have observed people using l2 loss to compare output map and gt heatmap (masked by segmentation map).
But I could not exactly understand your implemenation of first doing soft_argmax and den comparing the coords instead of maps ? isn't that too sensitive ?
Is the loss based on "comparing 2 maps" or are you subtracting the exact coordinates ?
In my experience I have observed people using l2 loss to compare output map and gt heatmap (masked by segmentation map).
But I could not exactly understand your implemenation of first doing soft_argmax and den comparing the coords instead of maps ? isn't that too sensitive ?