Perceptual Organization From a Bayesian Point of View

Speaker: Jacob Feldman , Rutgers University

Date: Friday, April 21, 2017

Time: 4:00 PM to 5:00 PM Note: all times are in the Eastern Time Zone

Public: Yes

Location: MIT Singleton Auditorium, 46-3002

Event Type:

Room Description:

Host: Prof. Tomaso Poggio

Contact: Kathleen Sullivan,

Relevant URL:

Speaker URL: None

Speaker Photo:

Reminders to:

Reminder Subject: TALK: Perceptual Organization From a Bayesian Point of View

Abstract: Perceptual organization is the process by which the visual system groups the visual image into distinct clusters or units. In this talk I'll sketch a Bayesian approach to grouping, formulating it as an inverse inference problem in which the goal it to estimate the organization that best explains the observed configuration of visual elements. We frame the problem as an instance of mixture estimation, in which the image configuration is assumed to have been generated by a set of distinct data-generating components or sources (``objects''), whose structure, locations, and number we seek to estimate. I'll show how the approach works in a variety of classic problems of perceptual organization, including clustering, contour integration, figure/ground estimation, shape representation, part decomposition, object detection, and shape similarity. Because the Bayesian framework unifies a diverse array of grouping rules under a single principle, namely maximization of the Bayesian posterior---or, equivalently, minimization of descriptive complexity---I'll argue that it provides a useful formalization of the somewhat vague Gestalt notion of Pr├Ągnanz (simplicity or "good form").

Joint work with Manish Singh, Erica Briscoe, Vicky Froyen, John Wilder and Seha Kim.

Research Areas:

Impact Areas:

See other events that are part of the Brains, Minds and Machines Seminar Series 2017.

Created by Kathleen Sullivan Email at Wednesday, April 19, 2017 at 5:54 PM.