Electrical Engineering-Systems Dept.

*** SEMINAR ***

Roy Josef Jevnisek

(M.Sc. student under the supervision of Dr. Shai Avidan)

On the subject:

Semi-Global Dense Matching

Semi Global Matching (SGM) is a method for dense stereo matching. It was originally formulated to work on two 2D images and we extend it to densely match two 3D volumes.

Unfortunately, SGM uses a simplified smoothness measure that is not suitable for the challenging 3D data we consider. Therefore we reformulate SGM to work with any smoothness measure. However, the time complexity of the algorithm grows quadratically with the number of labels, making it intractable for high dimensional data. To address this, we introduce an efficient implementation for a truncated L1 measure. Quantitative comparison between our method and a state-of-the-art alternative show comparable results.

Finally, we found that occlusions pose a major challenge to our algorithm and propose a modification to SGM that compensates for occlusions. Evaluation on the Middlebury data set shows major improvements.