Projects
Diversity and Uncertainty for Discrete Graphical Models
Inference for discrete graphical models (aka MRF/CRF/WCSP/...)
Solving MAPinference exactly if its LP
relaxation is good



Outoftheshelf combinatorial
solvers (like CPLEX) can solve really big energy minimization
instances, if you propelry apply them to a proper place in the model,
where their power is actually needed, where the relaxed solution is not
integer. Profit:
Global MAPOptimality by Shrinking the Combinatorial Search Area with Convex Relaxation NIPS2013 [bib] [pdf] [presentationpdf] [posterpdf] The code is available in OpenGM library (inference class CombiLP ) See also results of our solver CombiLP in the OpenGM benchmark. Work
in progress.

Partial optimality for MAPinference 


Metaalgorithm to turn
(approximate) inference solvers to provide a part of a globally optimal
(nonrelaxed!) solution of the MAPinference problem.
Partial Optimality by Pruning for MAPinference with General Graphical Models ArXiv 1410.6641 [bib][pdf]  Best student paper Award at CVPR 2014! A. Shekhovtsov, P. Swoboda, B. Savchynskyy Maximum Persistency via Iterative Relaxed Inference with Graphical Models CVPR 2015 [extended abstract] [bib][pdf with supplementary material] Check also [External project page] for a public code! Work in progress.

Local polytope (LP) relaxation for MAPinference problem. 


Efficient
estimation of feasibe primal solutions
for the relaxed MAPinference problem.
Profits:
B. Savchynskyy, S. Schmidt Getting Feasible Variable Estimates From Infeasible Ones: MRF Local Polytope Study. In Advanced Structured Prediction, MIT Press, 2014 [bib] [pdf][posterpdf] The code is available in OpenGM library (inference class PrimalLPBound ) 

Smoothing technique for solving LP relaxation of the MAPinference problem. Profits:
A Study of Nesterov's Scheme for Lagrangian Decomposition and MAP Labeling In CVPR 2011  oral presentation [bib][PDF with supplementary material][Presentation (pdf) at the Graphical Model Workshop in Kiev, Sep.Oct.2010 and at CVPR2011] The code is available in OpenGM library (inference class NesterovAcceleratedGradient) B. Savchynskyy, S. Schmidt, J. H. Kappes, C. Schnörr Efficient MRF Energy Minimization via Adaptive Diminishing Smoothing In UAI, 2012, pp. 746755. [bib] [PDF (revised version with appendix)][UAI Poster with additional comparison to ADLP algorithm] The code is available in OpenGM 2.1.0 library (inference class ADSal) See also results of our dual blockcoordinate ascent algorithm ADSal in the OpenGM benchmark. Journal paper coming soon.

There are also more new projects running... and a number of others, which I less involved to. See my publications for details.