Hi,

I'm using the free trial with 2 Kinects set up in an approximately 80 degree configuration.

My test motion is throwing a ball overhand (slowly) with my right arm. I've found that if I TPose facing camera 1 and then perform the throwing motion facing in the direction that bisects the angle between the 2 Kinects, I can record a motion that has no occlusion. If I face camera one, I get some occlusion of the right arm in it's backward position. But in general, the software does a reasonably good job capturing the throwing motion (including the feet).

What surprises me, however, is that I sometimes lose tracking on the left arm when it's in front of my body. Moreover, when that arm is extended somewhat away from the body, the tracking algorithm fails to automatically retrack the arm. In one instance, the tracking only found and retracked the arm at the end of the motion when it was almost fully extended downward. (I was beginning to walk back to my computer to stop the capture session.) Is this an inherent problem with the current version of the ipiSoft software?

I also find it surprising, that a real-time skeletal program such as Brekel seemed to have little problem tracking my arms when they were in front of my body (but occlusion was significant on the backward portion of the motion). I would expect that a post-processing solution would not struggle with this situation.

I'm looking to eliminate any cleanup on the basic motion. Can someone comment on whether there's a way to eliminate tracking problems when my arms are in front and relatively close to my body? And why doesn't the ipiSoft software do a better job of automatically retracking the arm when it moves to a position where it is clearly separated away from the body?

Thanks

Hi

Our software employs incremental tracking algorithm, where the tracking of the following frame is based on the pose in the previous frame. That is why it maybe hard for the algorithm to recover from tracking errors without manual intervention. Usually such errors are easily fixed by tracking backwards from the frame where tracking was restored (either automatically, or with Refit Pose). Currently, this is a limitation of our technology, which we hope to overcome eventually.

Brekel software employs Kinect SDK built-in realtime algorithm which is based on pose recognition using predefined database of poses. It recognizes each frame independently, and thus recovers from tracking errors automatically. The realtime and auto-recovery are essential for user input purpose which is a primary appliance of Kinect. But this comes at the cost of accuracy which is not always suitable for mocap.