Kinect Fusion Mesh deformed

Question

Hello,

I've recently acquired a Kinect v2 and I was experimenting with the Fusion tool for scanning objects. I've been using the Fusion Explorer from the 1.8 Developer toolkit for quite some time linked to a Xbox 360 Kinect and I've managed to pull off some pretty neat scans. So you can imagine I was stoked to try out the V2. My laptop is running it fine, everything checks out in the Kinect Configuration Verifier however, when I try to scan something, the information gathered at a much faster rate by the camera doesn't seem to be supported by the Fusion tool in the new SDK for the V2. 

Allow me to explain.

Usually when I wanted to scan something with 360 Kinect that required multiple passes and required me to move the Kinect sensor all around the object, I would set the Maximum Integration Weight slider to about 500, in order for the mesh to average over longer periods of time in order to have a more detailed reconstruction.

Well, in the new Fusion, even when setting the slider all the way up to 1000 "weight", the mesh updates absurdly fast and it averages it all wrong leading to the complete destruction of the mesh as soon as you go to move the camera. I've mounted the camera on a tripod and tried to slide it up and down. I had it on a level surface translating it horizontally. Nothing seems to work! It's really disappointing since the "static" scan resolution is a lot better but, if you cant scan in 3D... whats the point?

Does anyone know of a way to increase the weight even more. Is there a quick fix to this? By the way, the camera is running at a cool 30 FPS per second and when performing a "static" scan it does so beautifully.

I'm afraid I'll have to return the scanner :(

Thanks,

-Sebastian

Answer 1

What is the framerate of the system when scanning? How fast are you moving? Which version of the sample have you tried? Have you tried 3d builder windows store app as well? This video shows the technique of scanning.

https://www.youtube.com/watch?v=nXbDkJ7cRrA

Scanning has to be done slowly and can be effective if you remove a lot of the background areas so the scan focuses on the object being scanned, not the background points that may disappear. Because the sensor is based on time of flight, the physical reflective attributes can effect the data coming back. One of the things you can do is change the code and other integrate ever other frame or some other logic. As well, try to use the c++ version of the sample and see if that changes your results.

---

Carmine Sirignano - MSFT