The detection models used by Hotscreeen can only detect bounding boxes. It would be possible to train a model that detects the actual shape of each objects, but it requires to have a dataset of thousands of images labeled by hand. I will try to do it one day, but not before months.)