Please send something that will break this... it's just too good

reminds me a bit of a old v4 leap motion stuff , I can still see some issues but overall very good!

Yeah, still quite a bit of issues especially for truncated out of screen points, but that's issue which is being fixed as we speak. We have a bit easier time then you guys had back then as we do purely ego-centric and live in the era of great VLMs :)

Would have loved to have used this for my project how do I get access ? I have a ton more that use mediapipe

Very cool! We've got a bunch of @PlayCanvas examples showing how to leverage MediaPipe. Should we be developing MacroData ones too? How do I get started?

@playcanvas Hey, it's not aimed for general public. We use it for clients who need ego-centric 3d hand-tracking or detection

@playcanvas Aww, ok, fair enough. I'll stand down. 🫡

Looks awesome! What compute is it running on and does it run in realtime?

Is that really a fair comparison though? Mediapipe sucks.

Tell me what i should compare to then

plenty of SOTA monocular hand mesh/skeleton reconstruction models. you can look them up.

I see, can you be more precise (name them) ? I am happy to run it and see, I am really curious how they perform. The main reason why we decided to do custom stuff is simply because nothing was working.

@ch3njus 2 years old

@ch3njus Hamer will fail horribly

Nooice!!

this is great!

Is this open source 🙏

We don't plan to open source the underlying model unfortunately. But we would be happy to provide the tracking for you if interested

Sick! Which benchmarks are you evaluating on? (except your internal ones)

For detection/identification, we use purely our internal benchmark, I think everything OS is terrible. For 2d keypoints, hint is reasonable and probably only usable public benchmark (even reprojected hot3d/arctic suck!!) For 3d keypoints we have something internal

@DominiqueCAPaul How about EgoSuite? You find its 3d keypoint annotations good enough?

@DominiqueCAPaul Out of all the vendors it's probably most decent one, however having mo-cap data that look like in wild data is always the best :)