The Gaussian function used to create heatmap with Gaussian peak at the ground truth coordinates. And in the training loop the distance between ground truth training data and model predicted heatmap calculated using MSE loss function. MSE = (1/n) * Σ(yᵢ - ȳ)²
If yᵢ numerically equals (or near) to ȳ but the Gaussian peak are at different places in the heatmaps than the loss function gives false low loss value. What methods can I use to overcome this issue?
