A graph cut based adaptive structured light approach for real-time range acquisition

This paper describes a new algorithm that yields dense range maps in real-time. Reconstructions are based on a single frame structured light illumination. On-the-fly adaptation of the projection pattern renders the system more robust against scene variability. A continuous trade off between speed and quality is made. The correspondence problem is solved by using geometric pattern coding in combination with sparse color coding. Only local spatial and temporal continuity are assumed. This allows to construct a neighbor relationship within every frame and to track correspondences over time. All cues are integrated in one consistent labeling. This is achieved by reformulating the problem as a graph cut. Every cue is weighted based on its average consistency with the result within a small time window. Integration and weighting of additional cues is straightforward. The correctness of the range maps is not guaranteed, but an estimation of the uncertainty is provided for each part of the reconstruction. Our prototype is implemented using unmodified consumer hardware only. Frame rates vary between 10 and 25 fps dependent on scene complexity.