Model fitting and regularization (discrete optimization approach)

Yuri Boykov
University of Western Ontario

Many problems in low-level image analysis are based on model fitting (e.g. color models in the context of segmentation or geometric models in the context of multi-view reconstruction or stereo). This lecture discusses regularization-based formulations for multi-model fitting problems and the corresponding optimization algorithms. In particular, we discuss Minimum Description Length (MDL) principle motivated by information theory. The corresponding energy includes high-order sparsity term (a.k.a. label cost) that can be efficiently optimized by efficient UFL algorithms or generalizations of a-expansion (graph cuts). Keywords: compression, quantization, sparsity prior, label cost, MDL, graphs, K-means vs EM.

Presentation (PDF File)

Back to Graduate Summer School: Computer Vision