I recently conducted an experiment in which I had 2 kinect systems looking at a ladder. One kinect looking straight at the ladder which is a 7 foot standing ladder (if the actor was climbing the kinect would be looking at the actors back). I also had a dual kinect system at a around 150 degree angle with the kinects on the sides of the ladder. The experiment was to validate if the dual kinect system would improve accuracy. I thought it would considering when the single kinect looks straight on, the arms get occluded sometimes behind the arm and the body of the actor( since ladder becomes narrower near the top of the ladder and the arms go in). Also the dual kinect has the added vision of seeing both arms at all times so it should produce more accurate data. I extracted the 3D spacial coordinates of the joints in the single kinect video and dual kinect video and compared it with the coordinates of the joints retrieved from the VICON 3D LAB setup (my base case). The video was taken with all three systems at the same time and compared to reduce variables. The video was of an actor climbing the ladder 2 steps up and down 25 times. The interesting result though was that while the dual kinect had improved accuracy for the left and right elbows (joints I chose to compare), the dual kinect system proved to be more inaccurate for the left and right knee. Could anyone propose some explanation for the reason in the drop of accuracy for the legs? Thanks

It is very difficult to answer your question based on brief information you provided.

In particular, "more accurate" and "more inaccurate" without numerical values and detailed description of measurement method can be understood in a large number of totally different ways. Do you get more jitter? Or larger discrepancy between iPi and Vicon systems? Is it a constant discrepancy for the entire duration of the motion, or a variable discrepancy? How the discrepancy is affected by placement of markers for Vicon system?

You would have to publish a scientific paper to properly explain the results of your measurements.

When comparing mocap systems based on different technologies, it is also important to make sure that you correctly place markers on actor so that you compare "apples to apples" rather then "apples to oranges". For example, Vicon system has very good "nominal" accuracy in a sense that it can measure the location of marker very accurately. When marker is placed on actor clothes, it can move (relative to actor body) with clothes. So in terms of accuracy of measurement of angles between actor's bones Vicon system is not as accurate as one may think from the tech specs of Vicon cameras. Moreover, particular method of placing markers on actor can result in different skeletons and therefore different values for angles between bones.

One can expect that a Kinect-based mocap system is less accurate then Vicon system, for the simple reason that Vicon uses much more advanced and expensive cameras.

As for the difference in accuracy between single-Kinect and dual-Kinect configurations, it is hard to provide an explanation based on info you provided. It is possible that single-Kinect is actually more accurate for legs. It is also possible that there is a flaw somewhere in your measurement methodology.

Another question I have is the field of view with the single Kinect vs. the dual Kinect. When a performer moves far away from one Kinect and closer to the other, meaning sufficient depth data from one Kinect but not the other, I could not get any results( the skeleton looks like its flying). Do dual Kinects work only when a performer stays within the operating ranges of both cameras? I am asking whether the range of the performer is smaller for the dual Kinect as opposed to the single Kinect.

I always felt the space with dual Kinect was a little smaller myself, at least for maintaining that extra level of accuracy. I haven't specifically tested for this but it seems apparent after having worked with a lot of dual Kinect shots.

I also feel like the accuracy can be better with 90 degree placement--but this may depend on the action being captured. In Happy Box, for example, setting the two Kinects at 90 degrees facing the front of the performer was definitely better than 180. In fact, if you look at the shots in the film, shooting 180 obviously doesn't make much sense for this production. Now, if the characters are walking around and completely turning away from one side to face the other, then the 180 configuration makes more sense.

This is just my opinion of course.

G.

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

Watch a one minute excerpt on Vimeo now!