Manuel Werlberger



I will soon release a public version of my image processing library (heavily using CUDA). It will include variational denoising algorithms and of course algorithms to solve correspondence problems like stereo and optical flow. Check my GitHub page :octocat: for the upcoming release.


During my Ph.D. studies I contributed to various software projects. In our research group we try to combine all our work to platform independend and reusable software that is used within the institute, by other research groups and our industrial project partners. Besides the software projects that are presented in the following I am mainly using C/C++ with CUDA for productive software, and Python and Matlab for prototyp- ing and evaluations. In terms of operating systems we support Linux, MacOS and Windows whereas we use Linux as the main developement platform in our group.


Optical flow describes the apparent motion of objects in a moving scene. It is one of the fundamental low-level Computer Vision problems and used in various tasks in the field where motion information is utilized. With the beginning of my PhD studies I started to develop a C++ library for computing optical flow and throughout the last years a versatile framework evolved that offers the possibility of computing optical flow in a very fast and accurate manner. In terms of speed and quality we provide one of the leading products in the field and the amount of feedback, requests and questions we get shows that the library is widely used in the research community. In addition the software is used in various commercial products and by our research and industrial partners.


Based on the achievments of the cudatemplates project I started another open-source (LGPL) project that evolved to the basis library in our research group. The library is also used to incorporate our libaries into commercial products by our industrial project partners. The main intention of this library is again the simple usage of the GPU as computing device and especially the connection of classical C++ programming with Nvidia’s CUDA toolkit. One of the benefits of the library is its modularity combining different modules and interfaces to other important software packages like OpenCV, Qt, Matlab or IPP. More information can be found on the project’s website.


The interest on GPGPU programming and solving difficult computer vision problems on the GPU was the starting point of the developement of an templated C++ library that simplified the usage of modern graphic cards as computing device. The main developer of the open-source project (GPL) is Dr. Markus Grabner. I not developing any further extensions of the cudatemplates as our research group has switched completely towards the ImageUtilities. More information can be found on the project’s website.