Computer Science Also Takes Second Place in Optical Flow Competition at KIT Germany

04 October 2012


Besides stereo vision results, optical flow is another basic input for computer vision systems, describing the movement of objects in a recorded dynamic 3-dimensional scene. Stereo vision and motion analysis both involve analyzing corresponding pixels, in stereo vision between the left and the right view, and in motion analysis between pixels in frames recorded at different times.

The correspondence analysis in stereo vision is performed along the epipolar line, thus defining stereo vision basically as a 1-dimensional correspondence problem. In motion analysis, there is not such a restriction; corresponding pixels can be in any direction, thus defining optical flow analysis as a 2-dimensional correspondence problem. Accordingly, error rates are still much higher for top-performing algorithms in optical flow calculation than in stereo analysis.

Since the early 1980s, researchers worldwide aim at solving the optical flow problem with reasonable accuracy. This task is a subject of major research programs in several countries.
The Karlsruhe Institute of Technology, Germany, and the Toyota Technological Institute at Chicago, USA, provide on their benchmark website challenging real-world data for a competition on best performing optical flow algorithms; see the ranking of optical flow algorithms on the KIT website.

The listed top-performing optical flow algorithms are published at excellent conferences or the best journals in the field. And researchers of The University of Auckland did it again! They are currently not only listed second in the stereo analysis competition, but now also second in the optical flow competition.

The design and implementation of The University of Auckland's successful optical flow program was also done by Simon Hermann, PhD student in the Computer Science Department, Tamaki Innovation Campus, supervised by Professor Reinhard Klette. Simon and Reinhard will present their optical flow solution at a workshop at the Asian Computer Vision Conference (ACCV) this November in Korea. ACCV is one of the excellent conferences in the computer vision field.

For more information about this research, please contact:
Professor Reinhard Klette