Organizers
Nikos Komodakis, Pawan Kumar, Nikos Paragios, Ramin Zabih
Brief description
Several problems in computer vision, pattern recognition, medical imaging and signal processing can be formulated using the discrete graphical models framework. The two main issues faced by researchers when using graphical models are: (i) Learning: How to estimate the parameters of the model?; and (ii) Inference: How to find the best assignment for the variables of the model? In this tutorial we will discuss these two issues, starting from the basics and building up to the state of the art.
Course outline
1. Introduction [slides]
2. Inference in graphical models
2a. Inference I [slides]
- Graph-cuts, move-making algorithms
- Submodularity/regularity
- Psedo-boolean optimization, roof-duality, probing
- Graph-cuts for higher-order problems
2b. Inference II [slides1] [slides2]
- Belief propagation
- Linear programming relaxations
- Dual decomposition
- FastPD
- Demo session: FastPD library
[Matlab wrapper for FastPD]
[Matlab stereo matching demo]
[demo data]
[FastPD download site]
3. Learning of graphical models
3a. Learning I [slides]
- Introduction to learning of graphical models
- Maximum-likelihood learning, max-margin learning
- Subgradient methods and constraint generation methods for max-margin learning
- Efficient max-margin training of high-order models via dual decomposition
- Learning of high-order latent CRFs via dual decomposition
- Case study: Learning to cluster using high-order graphical models with latent variables
3b. Learning II [slides]
- Max-margin learning with latent variables
- Self-paced learning
- Self-paced multiple-kernel learning
- Max-margin min-entropy models
- Demo session: self-paced learning
[Self-paced learning web-site]