Time taken to detect a skeleton with NITE

[Oni]

Hi all. I was wondering what experiences people were having with the
NITe libraries from Primesense when it comes to detecting a skeleton?
On average, once the pose is assumed I'm getting around 5-10 seconds
in most cases. I've noticed that wearing a coat can affect the outcome
(makes sense) but also, it would appear that the lighting makes a
difference as well. We've run tests in crowded places, places with
strip-lights and what not and we've had varying results from around 5
seconds from pose to never (by pose i mean from the detection of the
calibration pose).

Im wondering if noise in the depth readings is the main factor and if
I should be looking at what makes the stream noisy or if its something
I've done wrong in my code. I was initially using some code derived
from Roxlu's source in Openframeworks but now, its much closer to the
OpenGL primesense code. Im wondering if anyone else has can share
their thoughts on this and what are the best conditions for a speedy
detection?
Cheers!
Ben

[Nicholas Teeple]

I'd also like to hear more thoughts on this. I just ran an installation and there were multiple instances of participants seeming like perfect calibration candidates, only to have them stand there for 20+ seconds at times.

The stream is extremely noisy based on the preview window from OSCeleton. I also had the thought that back-lighting someone with an IR emitter to better differentiate them might help, but I might be way off on that.

[Nicholas Teeple]

Here's a post from Sebastien over at the tryplex group

These are my results so far on what could have a negative influence when calibrating/using the kinect:

- Shiny/black/leather clothing materials
- Clothing that could 'deform' the natural body shape like long coats/baggy etc.
- Direct sunlight positioned towards the camera
- The wall/background materials. Concrete seems to work less.
- Small contrast between background/person
- Large spaces (no background or objects in the field of view of the kinect)
- Small children
- Camera positioning; it works better when the head is visible, check the live image with something like the kinect driver for qc if you're body is inside the camera's view. A distance from the camera of 2-4 meters is best.
- When implementing in a public installation marking the boundaries with gaffer-tape works well

Didn't have problems yet with tall people. Average it takes about 2-4 seconds depending on the conditions.
You'll have to experiment a bit with different spaces/lighting conditions and camera positioning.

[Oni]

Thats a nice summary! Cheers for that.

On May 9, 3:41 pm, Nicholas Teeple <nicho...@teepareep.com> wrote:
> Here's a post from Sebastien over at the tryplex group<https://groups.google.com/forum/#%21topic/tryplex-toolkit/ks7u2xzbrsY>