License Considerations in Skeletal Tracking

Hi,

I recently got my hands on an Intel D435i camera, and started trying out Nuitrack.
Currently, I am using the trial version, trying both in the Nuitrack application and the Unity extension.
During my tests, using the trial license I could see only less than ideal results on the skeletal tracking - the skeleton fluctuates a lot, and does not track legs and arms to a good extent, and fluctuates between identifying them or not identifying them.

Video files that were recorded: Files
Test1: Unity Demo (using the Nuitrack unitypackage).
Test 2: Nuitrack App Demo - Roughly the same input, going from A-Pose to T-Pose and moving the arms upward, then doing a full spin and changing the leg positions a bit.

I tried the settings that were shared on this forum correspondence and got roughly the same results.

The computer I am testing with is a Windows 11, Core I7-1165G7 with 16GB RAM, without a GPU (using the integrated intel graphics on board).

My question is whether the Pro license or other licenses have any improvement on the measurement?

Taking a look in the comparison chart I don’t see any difference between the trial and pro other than the time frame (3 minutes are suitable for only testing the skeletal tracking).

I would assume using the AI license should work however I don’t have a GPU, and the device I am planning on using which are embedded devices with limited capabilities (with an integrated NPU/GPU that measures 5 TOPS) so I don’t know if it will work smoothly.

As you can see the tracking of the edge of the arms on a simple T-Pose does not measure the hands correctly and the whole skeleton fluctuates. This is after conducting calibration on the camera using Intel documentation and trying out different parameters. Is there something I done wrong?

Hello @ofek

The AI skeleton runs on the CPU. You can test it using a trial license.

I would assume using the AI license should work however I don’t have a GPU, and the device I am planning on using which are embedded devices with limited capabilities

At the moment, AI only works on x64 platforms. What architecture is your embedded device based on?

Thank you for your response irakli,
I was planning on using an ARM based device (64 bit OS), is it not supported by NuiTrack?

Nontheless I found the issue which was not caused by NuiTrack, but rather the Intel Realsense camera. I was testing it in an office environment, and discovered that strong lights emitted by both LEDs and fluorescent lighting caused the weird detection, and has confirmed it using Intel Realsense Viewer and a lot of other threads by other people on that camera.

Hi @ofek

I was planning on using an ARM based device (64 bit OS), is it not supported by NuiTrack?

Yes, supported. Could you specify which device you plan to use?(e.g. Raspberry Pi 4/5, Rock Pi, jetson)

Yes, we will use a Rockchip RK3588 based device, most likely will go with the Orange Pi 5, or the Rock Pi 5B.

Does it require any package or a special license when using this configuration?

Hi @ofek, please take a look at our recent video for the expected behavior of Nuitrack AI specifically for D435i - tracking should be more stable than with “classical” engine given limited quality of depth data (on your video depth data is almost completely missing on arms).
Another possible takeaway from the video demo - compact Intel NUC-like (MiniPC) devices could be an alternative option to ARM-based embedded devices (RPi, RochPi, etc).
If your commercial project crucially depends on stable tracking with ARM-based device (like RockPi) + D435i - we could work out technical options for you (cc @leonid).