Keypoint visualization not working as in the original paper after training the model

[Skytingle]

I am using an online image after training the model to reproduce as in the paper below with a single image.

After using my own image, below is the result. Please is there anyone that has successfully done this after training with his own image. Please can anyone help. Thanks in advance.

Successfully training of the model

[silentchaostheory]

This seems like an issue of scaling, if you look at the labels on the x, y, and z access, the ground truth is a much smaller area. I'm not sure how the scale is conserved between the GT and the prediction.