Converted Coreml output model is not the same as the original model

[AnasArram]

Hi, I have converted HRNET model for pose estimations from pytorch to coreml successfully. But the output of the converted model is not as in pytorch model. I mean: the output of the model is the body joints keypoints, the issue in converted coreml is that the predicted keypoints are not exactly at the same position on human on the camera. (All keypoints are there but a little bit moved when you visualize them).

Any suggestion?