Event Scheduled for Nov 16, 2011
Event: BECAT Seminar Series - Multi-object estimation poses three questions: How many objects are in the scene? Where are there objects? and: Which object is which?
Location: ITE-Room 336
Time: 12:00 pm
Details of Event:
Abstract: Multi-object estimation poses three questions: How many objects are in the scene? Where are there objects? And: Which object is which? Let’s ignore – at least for now – the first, which is admittedly very important but here is not germane.
To estimate the location of (say) two objects from prior information and measurements is, for those of you who like to think in these terms, an optimization. The familiar metric is mean square error, and to minimize it answers both the second and third questions. But suppose that we are interested only in the second question. We can recognize this a relaxation of a constraint in our optimization; can the best metric improve?
In this talk it is shown that it can, and the result is (unfortunately) called MMOSPA. We find a rich area of label-free estimation, both static and dynamic. Versions of the trackers are shown to beat coalescence problems for closely spaced targets. And for those of you who like labels, it is shown that labeling can be re-inserted to the estimates, but in a way that is perhaps more honest than for labeled ones, as a posterior label probability.
Sponsored By: Booth Engineering Center for Advanced Technology
Pamphlet/Flyer: No Pamphlet/Flyer Available