Trainer should accept dataset with bbox only

[qgeissmann]

There are popular datasets with bbox only. Instead of either making masks or not using them, we could learn boxes on them and "mask" the mask losses. One can hypothesise that detection improves overall for close domains... to be tested