Research Groups

Portrait Carsten Rother

Carsten Rother

Machine Learning, Computer Vision and Discrete Optimization for Bioinformatics

Previous and Current Research

The newly launched Computer Vision Lab Dresden (CVLD) aims at developing novel theoretical concepts which are practically relevant. On the theoretical side we work on discrete optimization and machine learning techniques mainly for so-called structured models – also known as undirected Bayesian networks/graphical models or Markov Random Fields. Structured models give rise to a very generic and commonly used modelling language which find their application in many diverse fields. The basic idea is that elements, e.g. pixel, voxels, or particles, in 2D, 3D or high-dimensional spaces, do not make independent decisions but rather joint decisions. Typical practical questions are: “How likely is a certain pixel belonging to a cell or not”;  “What is the 6D motion of a 3D particle”; “What is the true pre-colour of a pixel prior to applying the image formation process”, or “What is the 3D Layout of an organize”. The input data can be 3D Microscopic data or videos from an RGBD camera. In the space of Bio Imaging and computational Biology, we closely work with various other groups such that of Gene Myers, Ivo Sbalzarini and Pavel Tomancak. The first, joint project has been a new technique for automatically matching an instance of a 3D C.Elegans recording with a statistical atlas of a 3D C.Elegans. For this we have developed the “active graph matching” technique, where we combined an active shape model with a sparse structured matching model, known as graph matching.

All nuclei of a known C.elegans atlas (top) is automatically matched against a new C.elegans instance (bottom) with the help of our new “active graph matching” approach.
Future Projects and Goals

The newly launched Computer Vision Lab Dresden (CVLD) aims at developing novel theoretical concepts which are practically relevant. On the theoretical side we work on discrete optimization and machine learning techniques mainly for so-called structured models – also known as undirected Bayesian networks/graphical models or Markov Random Fields. Structured models give rise to a very generic and commonly used modelling language which find their application in many diverse fields. The basic idea is that elements, e.g. pixel, voxels, or particles, in 2D, 3D or high-dimensional spaces, do not make independent decisions but rather joint decisions. Typical practical questions are: “How likely is a certain pixel belonging to a cell or not”;  “What is the 6D motion of a 3D particle”; “What is the true pre-colour of a pixel prior to applying the image formation process”, or “What is the 3D Layout of an organize”. The input data can be 3D Microscopic data or videos from an RGBD camera. In the space of Bio Imaging and computational Biology, we closely work with various other groups such that of Gene Myers, Ivo Sbalzarini and Pavel Tomancak. The first, joint project has been a new technique for automatically matching an instance of a 3D C.Elegans recording with a statistical atlas of a 3D C.Elegans. For this we have developed the “active graph matching” technique, where we combined an active shape model with a sparse structured matching model, known as graph matching.

Methodological and Technical Expertise
  • Building Probabilistic and deterministic structured models of the image formation process and organisms (e.g. Markov Random Fields, Conditional Random Fields, undirected Bayesians networks, and undirected graphical models)
  • Inference in structured models - discrete and continuous, probabilistic and deterministic optimization, e.g. OpenGM, higher order models, Dual Decomposition, etc.
  • Data-Driven Machine learning in particular for Structured Models, e.g. Random Forests, structSVM, etc.
  • Image matching (optical flow, tracking, graph matching)
  • Image segmentation (graph cut, interactive GrabCut, proposal generation)
Selected Publications

Dagmar Kainmueller, Florian Jug, Carsten Rother, and Gene Myers
Active Graph Matching for Automatic Joint Segmentation and Annotation of C. elegans
MICCAI 2014

Uwe Schmidt, Carsten Rother, Sebastian Nowozin, Jeremy Jancsary, and Stefan Roth
Discriminative Non-blind Deblurring
in 2013 Conference on Computer Vision and Pattern Recognition (CVPR 2013), IEEE Computer Society, 20 April 2013

Markov Random Fields for Vision and Image Processing
Edited by Andrew Blake, Pushmeet Kohli and Carsten Rother, MIT press 2011

Carsten Rother, Vladimir Kolmogorov, and Andrew Blake
GrabCut-Interactive Foreground Extraction using Iterated Graph Cuts
in ACM Transactions on Graphics (SIGGRAPH), August 2004

Oliver J. Woodford, Carsten Rother, and Vladimir Kolmogorov
A Global Perspective on MAP Inference for Low-Level Vision
ICCV, 2009

CV

since 2013
Full (W3) Professor at TU Dresden heading the Computer Vision Lab Dresden (CVLD)

2003–2013
Researcher at Microsoft Research Cambridge/UK

1999–2003
PhD Student Candidate at Royal Institute of Technology in Stockholm (KTH)/Sweden

1996–1997
Computer Science at the University of Edinburgh/UK

1992–1999
Computer Science at the University of Karlsruhe/Germany

Contact

Computer Vision Lab Dresden
Technische Universität Dresden
Nöthnitzer Straße 46
01187 Dresden

Homepage

Intranet