Training not working #15
Description
I tried to train the model with H-3 config but it is not working. I tried to use make_heatmaps on training outputs and it seems that the model is always guessing the keypoints in the black zone that is not part of the original Coco image (where the mask is 0). Here is an example.
I'm not sure, is the mask supposed to cover the entire original image? Or is it suppose to cover only the people on the image? Because I get this
