Silhouette-based 3D reconstruction

The 3D reconstruction of real world objects from several images is a current and exciting research branch in computer vision. At the same time, the demand for 3D models in the movie and game industry is increasingly growing. A simple and cheap way to reconstruct objects is using only one camera. Within the framework of my master's thesis at the RheinMain University, I've developed a hardware and software system which achieved this goal.

The hardware consists of a turntable on which an object might be placed, a stepper mitor that drives the turntable, a controller which triggers the motor and a (calibrated) camera which takes several images of the object.

From the video input of the camera the object first gets separated from the background using a thresholding method and then an initial 3D reconstruction (so called visual hull) is estimated using a voxel carving method. This reconstruction is then further refined using the texture of the object.

The obtained results are evaluated and the limits of the system are analyzed. A software interface allows the user to use the entire system to create a 3D reconstruction of an object in a simple and fast manner.

Squirrel voxelgrid
Squirrel interim result