June 21st, 2012 @ 2:30pm Vittorio Ferrari (ETH)

Two ways to learn an object detector without bounding-boxes

Abstract: To solve object class detection in complex, real-world images we need powerful visual learning techniques to acquire rich models capturing the diversity of the visual world. To learn complex models and scale up to a large number of classes, learning should require as little human supervision as possible. In this talk I present recent research on visual learning at the CALVIN group, with focus on learning object detectors from images or videos labeled only by which class they contain, but without information about their location in the image. In particular, I will present two recent works on (a) learning object detectors from web videos, where motion segmentation can help localising objects, and hence replacing bounding-box annotation; (b) transferring knowledge from ImageNet classes with available bounding-box annotation to related classes without annotation.

http://www.vision.ee.ethz.ch/~calvin/

seminars/seminaritems/2012-06-21.txt · Last modified: 2012/06/18 12:58 by silberman