Provide pre-trained model with a more suitable training set

[enkore]

This currently uses the "standard" pre-trained model, which a) uses 80 classes, out of which only very few are relevant, which means that even the "tiny" model actually has a lot of potential left to optimize only for the classes which are relevant. There are also likely better training datasets than COCO for this application.