Explanatory probabilistic models for temporal scene understanding
I will address some methodological topics relevant to TRIPODS research working groups 3 (images), 7 (Bayesian methods), and 8 (spatial and temporal data), using work on inferring what is going on in the world from image and video data as a running example. Bayesian approaches span a large group of group of methods that I will briefly mention, and then I will focus attention on problems where the representation has meaning outside that implied by probability distributions. Similarly, I will discuss briefly how image data can be processed to an advantage for classification (for example) without bothering with underlying semantics. I will then discuss the alternative that we use in our scene understanding work, which is to focus on the underlying representation of the world as explaining image data.
Pizza, Coffee, and Tea will be provided