I am trying to find an accurate and fast multi-person human pose estimation that I can train on with custom data. I have been searching for a little while and I may not be up-to-date on the newest techniques. I will start by posting what I have found and looked into (a little):
- Openpose: This is supposedly real-time (I assume on a GPU, 24fps?) and they provide training code
- Lightweight OpenPose: Runs in realtime >20fps confirmed, training code is provided
- mediapipe: runs in realtime > 20fps confirmed, training code is NOT provided
- posenet: No training code, can one even train tfjs models?
- movenet: Very fast but no way to train?
- hrnet & lightweight-hrnet: seem to be slow? Can anyone confirm? training code provided
- blazepose: haven't tried it yet, looks like tf implementation, but no discussion bout speed. Training code included
- alphapose: haven't tried it, but looks to run at 16fps (maybe faster?) but is only intended for research not commercial. Training script available.
- MoVnect: Looks new, and fast (haven't tested it) but looks like it uses student-teacher training.
What are some other human poses estimation models out there?
I care more about speed and training. I am thinking #2 is my best bet but it's a few years old. And not the most friendly to train Anything newer? Can anyone confirm or reject my findings?
