Building a Dedicated MOCAP machine


#1

I am trying to pull other posts regarding hardware together by asking a single question.

What would a computer setup dedicated to MOCAP look like?

If we were to go beyond the basic system requirements statement, and build a machine to do the best possible job with ni mate’s version of MOCAP - exclusively - what would that system use? The goal would be a square room with 3 sensors set up on the x, y, z axes. Each sensor capturing simultaneous - independent - data streams. Each stream would be captured on a separate computer and then transferred to a different system for integration in Blender. Please see my post titled: 3 Axis with Blender - for an idea about data integration.

[ total 4 computers - one for processing and rendering, 3 for data capture ]

Of course, the goal would be to reduce the cost of the data capture machines as much as possible. [ That is the impetus for the question. ]

Would this system be better designed to used the Kinnect and adapters or a different capture option?
What does the GPU need in terms of VRAM, speed etc? Can we use a card that does not cost $1000?

Again, I’m trying to go beyond the basic system requirements statement and discover what professional users of this tech are finding to be best for function and cost.

Thanks


#2

The requirements for a single sensor aren’t as bad as they could be. Instead of doing 4 computers you could go with 2 and add a PCIE3 expansion slot for USB3 (assuming Kinect 2), as the USB3 bandwidth on a single computer can be the real problem. With Kinect 2, multiple Kinects requires the usage of libfreenect2 as opposed to the official Microsoft SDK (change it from the preferences).

Kinect 2 will have some need for a GPU as some of the processing is done on the GPU. We have used Intel NUCs, Gigabyte BRIX and Intel ComputeSticks earlier to some success. The main issue with those more packaged computers tends to be the Intel GPU which typically may not support interprocess texture sharing. A non-expensive GPU should work as long it supports most of the OpenGL 3.2 extensions, including the texture sharing. You probably can get away with only a couple of gigs of memory as NI mate isn’t really memory intensive. You’ll want to have some single core performance for the CPU as some of the processing tasks in NI mate are single-threaded.

However, as per your other message, the problems you’ll have will end up being about the IR dot patterns interfering, and you’ll be needing a solution for that.