AI Mocap Without a Suit: How It Works

So you said another important thing and being like video AI solution, something that we are going to also tackle and discuss, right? Because right now there are different ways of getting motion into a 3D space.

And that's one of the key things.

But you also mentioned about training your own AI to do this rigging, right? Or to replace the inversive kinematics.

That's also something that Meta has, let's say, pushed.

Okay, look, now we have legs.

Yes.

So how is somehow Animotive using part of that technology? Is it not? Can you tell us a little bit how do you build that AI mock-up system?

Yeah, so we captured around 15 hours of high quality Vicon motion captured data.

So this essentially looks like a bunch of performers having to Walk slow, walk medium, walk fast, change direction lots to create a huge data set.

So then what we're able to do is...

You know, we don't actually also have to capture a data set of the three points in space.

We can simulate those points based off the tracked skeleton.

But what this essentially means is we can make very natural movements and assumptions of what the legs are doing, where the head is based off the head and controller positions in 3D space.

meta have done their own versions of this um but what we've often found is uh there's some cheats along the way uh so they have just released another suite of tools to developers in the past week which i haven't had the chance to fully dive into yet um but uh uh within the past year their solutions were you know they would stretch the length of bones to make sure foot planting landed and things like that um but obviously that doesn't if it doesn't look good with an animated character for animation uh it's not a good solution for us

So I was also curious about when you mentioned 15 hours of 3D recording, right? About someone wearing a suit, it means that you're not recording a video, you're actually collecting 3D data about different joints and skeletons and so on.

Is this, because I don't have a term of reference, right? I'm just curious, like when we talk about 15 hours, is that really like a massive data set? Is like, this is okay? Or like, I mean, I just missed the term of reference.

It's always the challenge in the AI research world is how big is enough for your data set.

And it really depends on the research project that you're trying to do.

So as a frame of reference, I think the best thing you can do is look for somebody who's done something close.

work with similar volumes of data is typically the approach.

I mean, if you've got the budget and you can afford to just go crazy at the data end, that's fine.

But if you've got to capture it yourself, you've got to kind of work out how much is enough.

That's exactly the point, right? It's not because, oh, we should all have massive amount of data.

You want the right amount of data for the job, right? Yes.

And