I'd like to save other users from the lost time and added frustration I caused myself by attempting to perform calibration of 2 Kinects based on a reference plane (or "board" or "3D plane").

Here's a brief description of the problem scenario...

• Sensor configuration
  
  • Two Kinects in the 60-90 deg. configuration
  • I tried with sensor normals intersecting at roughly 60, 70, 80, and 90 degrees
  • See documentation of "First Configuration" at, http://docs.ipisoft.com/User_Guide_for_ ... igurations
  Calibration method
  
  • Reference plane / board / 3D plane
  • Recorded calibration scenes with many good frames more than once per sensor configuration
  • Used a 3x4 foot rectangular piece of cardboard
  • I "pitched and yawed" the board as much as I could without introducing excessive "noise" in the board's edges
  • See documentation of this method here, http://wiki.ipisoft.com/User_Guide_for_ ... alibration

All of my calibrations attempts failed in a few different ways,

1. Horrible quality calibration with resulting relative camera pose visibly "OK"
2. Poor quality calibration with resulting relative camera pose visibly bad, but not incorrect
3. Excellent quality calibration with resulting camera pose incorrect

The last failure mode is most troubling to me since it seems to indicate a degenerate case of calibration output that is not properly guarded against by the documented calibration procedure nor the Mocap software. The calibrated pose of one of the cameras ends up being rotated 180 degrees from its physical configuration. The calibration algorithm isn't able to discriminate which side of the plane the cameras are looking at in this case.

I have had no issues at all with calibrating any configuration of Kinect sensors using the light-based method. These configurations include two sensors (both configurations, 60-90 and 180 degrees), three sensors, and four sensors. I have done all of this with both Kinect v1 and Kinect v2.

Perhaps the documentation should reflect this priority that users should place on preferring the light-based method. If this were the case, I would not have wasted a whole afternoon on performing my first calibration with your software.

Mark

PS ... Maybe I did something wrong and/or this is a known issue. If this is the case, please let me know so I can try again and roll the corrected procedure notes into my internal documentation!

Hi

Actually, the board calibration method is mentioned only in a single note in the docs. And this note also states clearly:
Using flashlight marker is recommended as more easy-to-use method.

I suppose you did not move a board back and forth much. That's why the software misplaces the camera after calibration.
Only slight rotations of the board are needed/allowed. What is important - that a board is moved in relatively large range, at least +/-2 feet.

Maybe less rotation and more translation of the board will help, thanks for your suggestion.

I do think that the documentation suggestion to use the light based method is too weak considering the frustration this has caused me.

Rather, I would have appreciated a disclaimer instructing me to spend 40 minutes to run to a store and purchase a mini mag lite rather than waste this time attempting to calibrate with the board method.

Thanks for your constructive response, vmaslov. Cheers!

vmaslov pretty much covered it but, yes, moving the board towards and away from the devices is important.

When I'm using two sensors at <90 on one side of the room, I step forward and backwards one or two steps. This works perfectly every time.

If you do a <180 setup, you will probably need to move the board an even greater distance. I find that if move the board to where it almost goes out of view to each sensor, Calibration is always good. If you don't move the board far enough, the software seems to assume both sensors are on the same side of the room.

I also rotate the board a little but you'll find that if you over rotate the board, the software cannot track the board. If you watch the Calibration process carefully, you can see it losing the board when you rotate too far. My guess is that at a certain angle, the rays start glancing off the board instead of being bounced back to the sensor, resulting in errors.

Which brings up another point: don't use a board that has a glossy, reflective surface. It's also good to avoid solid black surfaces. I use a solid white sheet of foam core board but it's probably even better to use a gray or a neutral value color board.

Don't give up. Once you get the moves down, it's actually pretty easy and predictable to get a good calibration.

_________________
Greenlaw
Artist/Partner - Little Green Dog | My Demo Reels (2013,) (2015,) (2017,) and (2019)

Watch a one minute excerpt on Vimeo now!