Overview Description:
I'm using libfreenect2 to collect RGB and Depth Data from Kinect2 devices on Linux (Ubuntu 14.04) and generate PCL point cloud based on that.
The steps are generally: 1) libfreenect2::SyncMultiFrameListener::waitForNewFrame() to get the RGB and Depth frames, 2) libfreenect2::Registration::apply() to align them, 3) loop through 512x424 matrix and call libfreenect2::Registration::getPointXYZRGB() to fill all the matrix elements.
According to the performance testing result, step 3# was the one taking most of time in the whole pipeline. I was thinking to use parallel programming (either CUDA with GPU or multiple thread on CPU) in Step 3# to improve the efficiency. However, considering libfreenect2 has already providing CUDA/OpenGL options of pipelines etc. , it might looks helpful if libfreenect2 could provide one more function along aside with getPointXYZRGB() using CUAD etc. to generate all the points, i.e. adding libfreenect2::Registration::getPointXYZRGB(const Frame* undistorted, const Frame* registered, float** depth, uint_8** color), where 'depth' point to a 3x512x424 array of float representing point (X, Y, Z) matrix and 'color' point to a 3x512x424 array of uinit_8 representing color for related point.
Would you please kindly share your comments/thoughts about this?
Thanks in advance!
Overview Description:
I'm using libfreenect2 to collect RGB and Depth Data from Kinect2 devices on Linux (Ubuntu 14.04) and generate PCL point cloud based on that.
The steps are generally: 1) libfreenect2::SyncMultiFrameListener::waitForNewFrame() to get the RGB and Depth frames, 2) libfreenect2::Registration::apply() to align them, 3) loop through 512x424 matrix and call libfreenect2::Registration::getPointXYZRGB() to fill all the matrix elements.
According to the performance testing result, step 3# was the one taking most of time in the whole pipeline. I was thinking to use parallel programming (either CUDA with GPU or multiple thread on CPU) in Step 3# to improve the efficiency. However, considering libfreenect2 has already providing CUDA/OpenGL options of pipelines etc. , it might looks helpful if libfreenect2 could provide one more function along aside with getPointXYZRGB() using CUAD etc. to generate all the points, i.e. adding libfreenect2::Registration::getPointXYZRGB(const Frame* undistorted, const Frame* registered, float** depth, uint_8** color), where 'depth' point to a 3x512x424 array of float representing point (X, Y, Z) matrix and 'color' point to a 3x512x424 array of uinit_8 representing color for related point.
Would you please kindly share your comments/thoughts about this?
Thanks in advance!