In the video, Mickrosoft researchers Minmin Gong and Xin Sun demonstrates that the raw data from Kinect's depth sensors is actually much rougher than as one might have imagined. To produce a smoother and "water-tight" model, they use a process called "Poisson Surface Reconstruction" to process the spatial points into more realistic surfaces.
To make the algorithm practical for a gaming platform like the XBOX 360 with limited computing power and memory amongst other game code and resources, they were able to optimize the algorithm to reduce 80% of triangles in the 3D mesh with minor change to the resulting shape which cut the process down from 20 seconds to just 2.
"The Kinect-based Object Digitization project enables the creation of 3-D objects based on just a couple of snapshots. Color/depth images of the front and the back of an object are registered quickly in two dimensions, and a GPU-based surface reconstruction results in a smooth 3-D model," reads the description of the video.
The researcher says,
"Working with Xbox Good Science team, we developed a Kinect based solution for scanning 3D object. Our solution only needs two snapshot of the object: a front side color/depth image of the object and a back side color/depth image of the object. A fast 2D registration algorithm is used to register the front side and back side together based on object silhouette.
After that, a GPU-based surface reconstruction algorithm is applied to reconstruct the smooth and water-tight surface from the input data. Our solution is memory efficient and highly optimized for Xbox 360. As a fundamental support function, it has been successfully used in different Kinect Fun Lab applications such as Googly Eyes and Build a Buddy."