Machines that See, Powered by Probability

Add to your list(s) Download to your calendar using vCal

• Professor Andrew Blake, Microsoft Research Cambridge
• Wednesday 04 December 2013, 16:00-17:00
• MR2.

If you have a question about this talk, please contact CCA.

Extraordinary industrial seminar

Machines with some kind of ability to see have become a reality in the last decade, and we see vision capabilities in cameras and photography, cars, graphics software and in the user interfaces to appliances. Such machines bring benefits to safety, consumer experiences, and healthcare.

The visible world is inherently ambiguous and uncertain, so estimation of physical properties by means of vision tends to rely on probabilistic methods. Introducing regularization into functionals used in optimal estimation already helps absorb noise in sensory data, but visual processing makes further demands on the mechanisms of probability. Prior distributions over shape can help signficantly to make estimators of shape more robust. Learned distributions for colour and texture are used to make estimators more discriminative. These ideas support inference by finding hypotheses for the contents of a scene that explain an image as fully as possible. More recently this explanatory approach has somewhat given way to powerful direct estimation methods, with parameters tuned using large training sets. Perhaps the most capable vision systems will come ultimately from some kind of fusion of the two approaches.

This talk is part of the Cambridge Centre for Analysis talks series.

This talk is included in these lists:

• All CMS events
• CMS Events
• Cambridge Centre for Analysis talks
• DPMMS info aggregator
• Hanchen DaDaDash
• Interested Talks
• MR2
• My seminars
• bld31

Note that ex-directory lists are not shown.