Ideas for NuiTrack improvement

I have 3 ideas for NuiTrack improvement:

  1. The Intel RealSense cameras use for depth generation image analysis from 2 stereoscopic camera.
    Maybe you can add depth map generation with any pair of cameras (the algorithm is in OpenCV GPU accelerated for sure) so NuiTrack will work with a pair of any 2 identical cameras properly aligned ?

  2. Pseudo depth -map can be generated from a single camera based on pixel motion analysis. There already exist MOCAP products based on this idea (wrnch). Any chance to implement single camera MOCAP ?

  3. The l idea is to produce multiple standard camera more precise MOCAP

Your product and SDK look good so I believe it can be worth to extend its compatibility with any camera HW …

Thanks for your thoughts, Nuitrack strategically will work with any standard RGB(D) hardware. The road map includes new value added software features, and algorithm improvements based on AI, such as 3D face recognition, finger tracking, object detection and location awareness, as well as RGB skeletal tracking. we do not plan to be involved on a hardware level, there are plenty of companies who are experts on this :slight_smile:

I do not want to force you to do hardware R&D. The idea is to extend the support from RGB(D) cameras to multiple RGB cameras + D extraction from stereoscopic disparity(or just simple multiple RGB cameras) . This will give you super wide input HW coverage. I am just testing very intensively your product with all supported sensors +I had done head to head comparison to wrnch RGB tracking. Your product works OK but it need to work ‘‘a bit’’ better to be useful for more application. I do not see any other way for improvement than tracking based on multiple sensors data simultaneous processing.

Thanks again for your thoughts, btw, what is your conclusion on wrnch?
RGB body tracking is on our short term road map.
syncing several cameras we consider “application” level, not a SDK feature.

wrnch does not work with depth = it have issues to identify the skeleton-sensor distance and such distance change (user forward and backward motion according camera). In this direction the use of RGB(D) cameras have an advantage. General advantage of wrnch is it works with any camera (but it is not able to utilize higher camera resolution for improvement). The results (if I do not count the Z-coordinate issue) of wrnch and nuitrack are comparable in a slight favor for nuitrack.
According syncing more camera - I do not speak about HW synchronization. Cameras working and 60 Hz can give 1/120 sec sync (based on time-stamps) without HW synchronization.
Do you thing such a precision is not sufficent for MOCAP and eventual depth map construction form stereoscopic disparity map ?

Actually, this is what Intel RealSense is doing, but despite great efforts, it seems quite tough to get good results…