High-quality Stereo Matching Strategies for Random Speckle Pattern Stereo
The thesis should implement and evaluate stereo matching algorithms that are optimized for active random dot speckle projection texture. The thesis should especially focus on the effects of slanted surfaces and the resulting projective distortions of corresponding image patches observing the speckle projection. Completeness of the resulting depths maps in the presence of projective distortions is of particular interest. The problem can be addressed by optimizing the data term of stereo algorithms for this particular configuration and also by applying appropriate regularization and optimization techniques to get efficient solutions for different formulations of the problem.
A very efficient and low cost random dot speckle projector can be found in the popular Kinect1 depth sensor. The projector can be combined with a classical stereo camera rig. The added projected texture extends the applicability of stereo matching methods greatly and reduces the need for the presence of appropriate texture in the scene. The depth maps of the Kinect itself contain complex geometric distortions and can hardly be used to measure scene geometry accurately. This can be fixed by combining the Kinect projector with a custom stereo setup, that can be fully controlled and calibrated. The thesis will be conducted in the scope of the COMET-project Vision+ - Integrating visual information with independent knowledge (http://www.comet-visionplus.at).
C/C++, computer vision, 3D reconstruction