The Kapnetix team visits eNStudios in Poland

It's not very often that professional motion capture studios get clients who demand multiple errors in every shoot. But that's exactly what happened when Kapnetix visited eNStudios.

First, a huge thank you to the team at eNStudios. It was one of the most well organised and professionally run production houses that we have ever been to. Definitely check them out.

One of our primary focuses at Kapnetix is to collect as much animation data as possible to continue to improve our motion capture cleanup models, and to make them work on as many different animation styles as possible.

So we spent 20+ hours recording animations in eNStudios' impressive OptiTrack Prime X 41 studio, but there was a twist. We purposefully generated errors in each movement sequence to replicate the most common types of errors that mocap post-production teams have to deal with every single day.

And even though we caused these errors on purpose, the eNStudios team still spent hours and hours correcting the animations in post. The result is pairs of raw, error-filled animations along with their professionally edited versions. These are invaluable to training our models to fix the errors themselves, without an animator spending hours working frame by frame.

One common category of errors that we wanted to emulate is blocked or missing markers. This happens naturally in many different ways, but we decided to force these errors on purpose by covering random markers with black tape. For that shoot, eNStudios fixed it manually. That data is what teaches Kapnetix to do it automatically.

The next category of errors that we simulated was repositioning the mocap actors to the far edges of the capture space, out of range of the cameras on the far wall. Again, this is a common error that production teams work very hard to avoid, and fixing that many undetected markers by hand is a long day's work.

We plan to do more sessions with eNStudios to keep building up our training data. This capture is also the reason we can say that your takes never train our models: we own this data outright, because we commissioned it.