Dlib: Multi-class loss_mmod

[reunanen]

@davisking, thanks again for this awesome library.

You write here:

I haven't made the loss_mmod multi-class yet. It's a minor change though
and will be in dlib in the near future though :)

I guess this does not exist yet? At least couldn't find anything.

I'd be happy to try to contribute, however it seems you already have a clear idea how it should be done, so I guess I'd be better off waiting for your implementation?

[davisking]

I haven't done it yet. I'm working on the multi-aspect ratio update at the moment. Which is itself not that complicated. I'm fairly confident the code is right, but I haven't pushed it to github yet since I'm not done testing it. And testing DNN tooling takes forever because training is so slow.

Anyway, I'll push the multi-aspect ratio mmod in a week or two at the most unless I find something wrong with it, which seems unlikely at this point. Then I'll update it for multi-label support after that, which is a lot simpler and shouldn't require as much tedious testing.

[davisking]

As a partial update, I've pushed the multi-aspect ratio code to github a bit ago. It seems to be working great. I'm currently creating a vehicle detection example program with it, which is almost complete. Once that's ready I'll tag dlib 19.5 :)

Then right after that I'll add in the multi-label support.

[reunanen]

The multi-label support is now in master, and seems to work great. Thanks a lot, Davis!