We also bought the Nyko zoom module which hopefully will let the kinect work at shorter distances (good for exhibition environment and maybe for this 3D reconstruction experiment). They promise 40% zoom ie: 4ft effective distance for recognizing human skeleton. Kinect has an effective sensor range of 60cm (afaik)
I followed the instructions on Reconstructme.net and installed all the stuff they asked prior to launching the application. All went smoothly. I was happy to discover that the application is written using OpenCL which uses the power of the graphics card processors. It is like CUDA but all the 3 graphics chips manufacturers agreed to a common SDK. My Toshiba R705 laptop has an Intel graphics card for which there is OpenCL support. While it is not powerful enough to do continuous real-time 3D reconstruction, it should be ok for a basic taster of the technology.
Here's the challenge scene of a cluttered table top to be 3D reconstructed:
The CMD console where I ran the command line which opened the window showing the video from kinect plus the depth-map. Once I press 'p', the software tries to reconstruct the surfaces in real-time
Then it saves the surface into an STL mesh file.
Here is the STL mesh seen inside Blender. All looks well!
Now for some "poisson filtering" to reduce the noise and cover holes using Meshlab. However, there are way too many gaps in this 3D reconstructed scene. I will have to re-do the scene reconstruction using a more powerful computer and ensure the kinect sees the scene from all positions to avoid holes and gaps like this.
Here is a video showing what ReconstructMe developers promise to be a real-time 3D reconstruction of the scene, when used with a decent graphics card and cpu:
phpBB [media]
Statistics: Posted by limor — Tue Feb 28, 2012 12:24 am
]]>