Revision history [back]

I had the same problem. Taking a point-cloud approach is one way of doing it. I was trying to do this purely in Python, so that wouldn't work for me.

For the problem where the kinect rounded the depth values, that was easily fixed. We discovered that, when we converted to OpenCv, we used the command

bridge.imgmsg_to_cv(data, "mono8")

where bridge was an OpenCv bridge. However, this is looking for an 8 bit integer, and the camera produces a 32-bit floating point number, so this command should be:

bridge.imgmsg_to_cv(data, "32FC1")