Gathering some SOTA hand mocap models. Code under constructions. Eventually we will make a standard, ready for production version to do hand mocap along with body.
Here is the result.
More results are coming.
We provide 2 kinds of hand mocap model:
- With a hand box detector: more accurate;
- Without hand box detector, but need bodymocap, using bodymocap result to get hand box, less accurate but more pratical.
Except ohamo, other folders are vendor models with our modifications. **ohamo** is our OpenHandMocap System.
ohamo aiming at providing a fast, accurate, out-of-box 3D hand mocap system with training code. (ohamo itself is fully deployed running via ONNX doesn't need pytorch.)
ohamo is our pipeline, you don't need any prerequests before running, just install some package that you might don't have:
- alfred-py
Then, just run:
python main.py -i ~/Videoes/a.mp4
You will get a result on hand mocap.
Be note: Our hand detection model running realtime even on CPU, it's inference is on ONNXRuntime on CPU by default. For further accerlate please using TensorRT.
frankhand:
frankhand is a simpified version from frankmocap, we edited the pipeline to a hand detector free style, you can get hand without using hand detector;
mobrecon:
this is a decent hand mocap model, but code like a shit and messy. I simplifed it and make it more easy to inference.
- We will release a advanced hand detector model for easily setup hand detector;
- We will integrate with body mocap in future;
- torch
- open3d
- alfred-py
models supported:
- minimalhand;
- mobrecon;