BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Boosting in Location Space - Damian Eads\, University of Californi
 a Santa Cruz
DTSTART:20110519T200000Z
DTEND:20110519T210000Z
UID:TALK31498@talks.cam.ac.uk
CONTACT:Edward Rosten
DESCRIPTION:Computer-based object detection promises to vastly change our 
 lives.\nRobots will be able to map their environment and make sense of the
 \nworld. Scientists will have a new pair of eyes to sift through\nterabyte
 s of images of molecules\, proteins\, waterways\, and galaxies to\ndiscove
 r novel science. Most approaches to object detection use\ngeneral purpose 
 machine learning algorithms that optimize a\nnon-spatial objective. The cl
 assic sliding window approach predicts\nthe presence or absence of an obje
 ct in every window of an image\; this\noptimizes the classifier for detect
 ion but not for localization. In\ncontrast\, this thesis introduces a new 
 boosting algorithm\n(LocationBoost) that operates over an entire image at 
 all times during\ntraining\, directly predicts object locations\, and mini
 mizes a spatial\nloss function that is strongly motivated by object detect
 ion.\n\n\n\nThe research of this dissertation led to seven major contribut
 ions in\nobject detection. First\, since object detection is ill-posed\, a
  universal evaluation metric cannot be meaningful for all recognition task
 s.\nInstead\, we clearly define three different problems in small object\n
 detection and devise metrics well-matched to them. Second\, we\nintroduce 
 generative grammars to combine primitive image features into\ncomposite fe
 atures. Composite features are more informative and lead\nto more accurate
  object detection. Third\, AdaBoost is fragile in the\npresence of noisy a
 nd ambiguous training data but we made spatially\nexploitative adaptations
  to the learning algorithm to greatly improve\nlearning stability. Fourth\
 , a radically different boosting algorithm\n(LocationBoost) is proposed th
 at directly locates centers of small\nobjects\, bypassing the need for bou
 nding boxes. Instead of boosting\nclassifiers that predict whether or not 
 a patch contains an object\,\nour new approach boosts object detectors tha
 t produce a list of\npredicted object locations. LocationBoost uses a new 
 spatial loss\nfunction reflecting the intuition that large areas of backgr
 ound are\nuninteresting and not worth spending computational effort on. Fi
 fth\,\nLocationBoost is extended so it can predict bounding boxes that\nen
 close objects. Sixth\, a multi-scale variant of LocationBoost is\nproposed
  to enable the detection of both large and small objects in\nthe same imag
 e. In this variant\, we show how the structure of\nmulti-scale detection c
 an be exploited to greatly speed up training\nand detection. Seventh\, we 
 propose a new primitive image feature based\non FAST corner detection that
  enables real-time object detection.\n\n\n
LOCATION:Department of Engineering\, Board Room (video link to UCSC)
END:VEVENT
END:VCALENDAR