High-level computer vision tasks such as scene understanding and object recognition are nowadays routinely addressed with sophisticated statistical models, with limited success. I discuss what distinguishes these computer vision problems from other machine learning problems and how this poses unique challenges in model building.

One promising way to build better models is in enriching the model structure: the use of latent variables, hierarchical and deep architectures, higher-order interactions, and structure learning. To achieve this, while allowing for efficient estimation and inference, I first describe the behaviour of commonly used estimators for parameters of random field models and highlight their typical failure modes.

Emerging from these insights I propose a novel discrete random field model applicable to a large number of computer vision tasks. The model is conditionally specified, non-parametric, and able to represent complex label interactions, yet it can be trained from hundreds of images in minutes on a single machine.

AG 1, AG 3, AG 5, SWS, AG 2, AG 4, RG1, MMCI   Info

MPI Audience

English

45 Minutes

Saarbrücken

Peter Gehler

--email hidden

passcode not visible

logged in users only

machine learning; computer vision; probabilistic models