Vision and Image Processing (Former Research Group)

The Independent Research Group "Vision and Image Processing", headed by Andrés Bruhn, was part of the Cluster of Excellence MMCI from October 2010 until March 2012.

Vision and Research Strategy

We perform research on mathematically well-founded methods in computer vision and image processing based on variational methods and partial differential equations. Included in our focus are all types of correspondence and matching problems as well as their application in the context of image analysis.

We are interested in all aspects of these techniques, including mathematical modeling, well-posedness analysis, efficient algorithms for sequential and parallel architectures, as well as their application in real-world scenarios.

Composition of Group

The Vision and Image Processing Group headed by Andres Bruhn was founded in October 2010 in the Cluster of Excellence. The group consisted of two PhD students, Levi Valgaerts and Sebastian Volz. In addition, two student assistants, Ilya Goncharov and Stefan Densow, are employed by the group.

Research Topics and Achievements

Motion Estimation

Being of central importance for scene understanding, the estimation of motion information from a sequence of images is one of the key problems in computer vision. In this project, we aim to construct novel techniques for motion estimation that use a wide variety of priors that are self-adapting to the underlying scene. Apart from the design of such priors based on geometry, motion and scene structure, this project is aimed at developing novel methods to automatically combine them in an optimal manner.

Stereo and Scene Flow

Closely related to motion estimation is the recovery of the 3-D information of a scene from images taken from two or more viewpoints. If combined with a temporal evolution in terms of stereo or multi-view sequences, the estimation of the true 3-D motion of objects (scene flow) together with their surface (scene structure) poses a challenge of increasing importance. This project is aimed at modelling such techniques for the case of handheld cameras without external calibration. First results for a method that jointly estimates scene structure, scene geometry and scene flow from stereo sequences have been presented at ECCV 2010.

Image and Video Processing

A further research topic is the application of vision techniques for image and video processing. Typical examples are HDR imaging, superresolution, deinterlacing, framerate conversion, re-timing and panorama stitching. In all cases, the registration of images plays a major role. First results from this project, describing a method for the joint enhancement of the spatial and tonal resolution of images from handheld cameras, will be presented at EUROGRAPHICS 2011.

Efficient Algorithms

This project is meant to complement our modelling efforts in the field of computer vision and image processing by the development of efficient numerical schemes for sequential and parallel architectures. These methods include hierarchical techniques, such as unidirectional and bidirectional multigrid methods, as well as operator factorization approaches, such as our recent work on fast explicit diffusion (FED). A paper describing the latter approach has been awarded the Main Prize (best paper award) at the DAGM 2010.
Moreover, we have ported our algorithms to massively parallel architectures. Here, we focus particularly on the CUDA framework available for recent GPUs. In this context, we generalized our FED scheme to be applicable to the task of motion estimation. First results for one of our high accuracy optical flow methods (Complementary Optic Flow) on the GPU were presented at the CVGPU workshop at ECCV 2010.