How to impose a Loss to prevent Outliers (Unsupervised Landmark Detection)

Question

I am trying to predict full body landmarks in an unsupervised fashion. I am currently training on the Human3.6M Dataset and the results look as follows:

As can be