Stochastic Systems Group
		Home Group Members Publications Research Alumni Mailing List Demos Mission Statement Programs Seminars Library

Context may play a crucial role in detecting and localizing multiple object categories in a cluterred scene. For example, a monitor, a keyboard, and a mouse co-occur often in an office scene, and they tend to appear at specific relative locations to one another. We propose a hierarchical tree model to encode contextual information: instead of capturing every pairwise relationship directly, we introduce hidden variables (e.g., desktop) as common factors for related objects (e.g., monitor, keyboard, mouse). A tree hierarchy can represent co-occurrences and spatial relationships among hundreds of object categories in a compact way. We show preliminary results of a tree learned from a set of labeled examples.

Joint work with Joseph Lim, Antonio Torralba, and Alan Willsky.

Problems with this site should be emailed to jonesb@mit.edu