Dr. Boris Flach
Center For Machine Perception, Department of Cybernetics
Faculty of Electrical Engineering, Czech Technical University

121 35 Prague 2, Karlovo namesti 13, Czech Republic
Fax: +420 22435 7385, Phone: +420 22435 5776
E-mail: flachbor@cmp.felk.cvut.cz

My research activities in Structural Pattern Recognition and Computer Vision can be broadly grouped under the headings

  • Constraint Satisfaction, efficiently solvable subclasses
  • (Max,+)-problems on graphs and hypergraphs
  • Markov Random Fields and Boltzmann machines
  • Learning for probabilistic structural models (supervised and unsupervised)
You may also want to look on some recent applications:

Unsupervised (parameter) learning for MRFs on bipartite graphs. We consider a modified EM algorithm for unsupervised learning of Markov random fields defined on bipartite graphs. The main idea is to replace the intractable maximum likelihood estimator in the M-step by a tractable pseudo-likelihood estimator. The modified EM algorithm becomes tractable for MRFs on bipartite graphs and is competitive with persistent contrastive divergence - a popular method used for unsupervised learning in the context of deep learning. We apply the approach for learning a-priori shape models in the context of image segmentation.

cell_segmentation     chest-radiograph lung_segmentation

  • Boris Flach and Tomas Sixta,
    Unsupervised (parameter) learning for MRFs on bipartite graphs,
    British Machine Vision Conference, 2013, (pdf 1.5MB)

Modelling composite shapes by Gibbs Random Fields with pairwise interactions. We analyse the following question. Do GRFs with pairwise interactions have enough expressive power to model (1) spatial relations between segments, (2) simple shapes and (3) composite shapes?
original_tree_image segmented_tree_image    original_catman_image segmented_catman_image

  • Boris Flach and Dmitrij Schlesinger,
    Modelling composite shapes by Gibbs Random Fields.
    CVPR 2011, pdf 2.4MB
  • Boris Flach and Dmitrij Schlesinger,
    Modelling Distributed Shape Priors by Gibbs Random Fields of Second Order,
    Control Systems and Computers, (2) 2011, pp. 14-24 arXiv:1107.2807

Segmentation with shape and shading. We propose a probabilistic segmentation scheme, which is widely applicable to some extent. Besides the segmentation itself our model incorporates object specific shape priors and shading.
original_lady_image segmented_lady_image