Quantcast
Channel:
Viewing all articles
Browse latest Browse all 46

Hi Anatol D.

0
0

Hi Anatol D.

The reason I quoted the announcement part "At this time, we invite highly-skilled.. developers" was only to explain that *this* library has a minimal amount of documentation, samples, and no provided plugins for commonly used tools (such as Unity or Maya).  Also, the label 'experimental', and the fact that this is not part of the official SDK, was meant to manage expectations.  There are certainly many ways the technology in its current form is missing some important pieces (object interference, better recovery, two hands in close proximity, etc).  I certainly never meant to imply anything about anyone's skill level.  On the contrary, i would assume that anyone on this forum thread would have to be a fairly hardcore developer who knows motion capture, skeletal animation, etc.

I'll see what I can do about opening up the hand model data.  You are correct that the proportions do vary more than the limited scaling factors provided.  For example, males tend to have longer ring fingers than females.  Some people have fingers that are longer than their palms.  While tweaking the model does have some improvements, there is a point where the noise in the depth data itself surpasses the discrepancy in the model.  Also, The code in this library is limited to convex bone pieces, which limits how good of a fit there can be given the hand has some small regions of soft tissue (skinned like behavior) and concavity in the palm.  So while improvements are required, it will have to be more than just improving the tracking model.  

A camera's depth data cannot see that a ball is spinning.  Similarly the system will be challenged when trying to track a clenched fist being rotated.  A suggested path to improvements is to include some more traditional image processing and computer vision techniques - use the RGB or IR data to detect internal edges and features and track known markers.  These are areas where I begin to defer to other experts.  My background is more in rigid-body dynamics and 3d graphics.  

I'm not the spokesperson for the hardware or the official SDK.  Unfortunately, this forum might not be followed by the individuals in the know. You do have some good questions.  Please do ask, on the other forums, all your questions regarding hardware and software roadmap, as well as mentioning your developer needs and technology requests/expectations.   

Additionally, I can inquire internally about what can be done in the meantime about providing the requested improvements.  The original model came from a 3dsmax file marked up with additional properties to specify physical constraints.  It was exported to a custom xml format.   The ascii for the xml model file is simply an ascii substring inside the .dll file.  This was only done to make things more convenient - so developers would only have to copy one file (instead of two) into their own project's bin directory.  There was never any intention to hide the model data - its just 3d data that can be queried from the API.  Unfortunately this mesh data used for tracking cannot be modified in code.  (casting the vertices pointer the api provides to non-const doesn't help, the internal system uses a separate copy that is offset to each bone's center of mass)    The only effective way is to modify the input geometry.  I just tested editing the xml model description within the dll's binary directly (changed palm position z value from ~2.42 to 3.00 to add a gap between the wrist and the palm) and yes, it looks like the xml internal to the dll can be tweaked.  Probably ok as long as you dont change the filesize.  Although it might not be easy to understand parts of the file format. For example, the joint limit parameters aren't explained.  Note that I am certainly not recommending this approach.  It is not easy, nor do I suspect it would provide all the improvements that you would hope to achieve.  

 


Viewing all articles
Browse latest Browse all 46

Latest Images

Trending Articles





Latest Images