First of all, thanks for torch implementation! Finally i managed to run the model in a basketball environment. There was a few issue (in my case), but the overall result is promising.

• if you are using the labellingTool.py and you don't have w, h columns than the default value will be 50 pixel (sigma)
  https://github.com/mareksubocz/TrackNet/blob/12379a53da4ae343f064690686fd4bdd16da4283/dataset.py#L143C1-L144C43
  and because of this part, the heatmap is wrong (i just deleted it).

• for me the AdaptiveWingLoss was the right choice. Your constructed loss resulted a wrong anomaly in the training process (there was a weird, consistent blob in the predicted heatmap). I don't know why, I haven't checked it yet.

• I don't know why, but your predict.py did not work for me. I trained the model with images and tried to test with video files, but somehow the prediction was really bad. I rewrote that part and i read the images as you did it in the ImagesDataset class.