Gissu Valentina Naghavi, Huy Khanh Ha, Jan Frederick Eick, Nhu Quang Dang
Mariya Kaisheva, Adrian Kreskowski, Prof. Dr. Volker Rodehorst
Computer Science and Media (Bachelor of Science (B.Sc.)),
Computer Science and Media (english) (Master of Science (M.Sc.))
Type of project presentation
Summer semester 2019
- Bauhausstraße 11
attractive to children
One of the essential challenges in the field of computer vision is the estimation of 3D positions from images. The underlying task for the 3D position estimation is to find pixels showing the same content in different images. The process of establishing pixel correspondences is generally referred to as stereo matching. The output of stereo matching, which stores pixel differences between a pair of images, is called disparity map.
Despite being an active research area for decades, high-quality stereo matching remains an open challenge to this day. Although algorithms that provide highly accurate disparity maps exist, they are usually computationally expensive and therefore not well-suited for real-time applications. Faster algorithms, on the other hand, rely on local estimations and therefore tend to reduce the quality of the disparity maps.
In the “Real-time Stereo Matching” research project, we focus on finding a good trade-off between speed and accuracy for enabling live 3D reconstructions. For this sake, we implement and explore the extensibility of stereo matching algorithms which proved to be both efficient and of reasonable quality. To leverage the full potential of such algorithms, pixel correspondences are usually computed in a massively parallel manner on the GPU.
During this Summaery, we will demonstrate our work in progress application which performs live 3D reconstruction from stereo cameras using our own GPU-accelerated real-time stereo matching algorithms.