BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Learning Task-specific Object Location Predictors with Boosting an
 d Grammar-guided Feature Extraction - Damian Eads\, Univeristy of Californ
 ia at Santa Cruz
DTSTART:20090922T120000Z
DTEND:20090922T130000Z
UID:TALK20042@talks.cam.ac.uk
CONTACT:Edward Rosten
DESCRIPTION:\nBeamer is a new system for unstructured object detection: fr
 om an\ninput image\, it emits a list of (x\,y) pairs\, which are the predi
 cted\nlocations of objects. The system shows excellent results in the\npre
 sence of noisy and ambiguous greyscale aerial imagery\, and I\ndescribe ke
 y elements of the approach to achieve these results.\n\nFirst\, Beamer inc
 orporates domain expertise by generating a broad set\nof useful features w
 ith a stochastic generative grammar\, which also\neliminates many unhelpfu
 l features. A pixel classification is\nrepresented with a weighted linear 
 combination of these features\,\nwhich is learned with the popular boostin
 g algorithm AdaBoost. Second\,\nto improve robustness to label and data no
 ise without impeding\nlocalization performance\, we soften the learning go
 al of correctly\nclassifying every pixel in an object by optimizing a post
 -processing\nfilter applied to weak pixel classifications\, which makes us
 e of\nconfidence-rated AdaBoost. Since a purely pixel-based approach to\nl
 earning is limited for predicting locations\, the final detector is\noptim
 ized with (x\,y) location-space criteria.\n\nGrammar features are broadly 
 applicable to many computer vision\nproblems so I will give a mini-tutoria
 l on how to use them in your own\ncomputer vision research. GGFE (http://g
 gfe.googlecode.com) is a new\nopen source library written in Python for th
 is purpose. I\ndescribe the basics of generative grammars\, generating pro
 grams from\nthem\, and finally how to incorporate image processing algorit
 hms into\na grammar.\n\nFinally\, I conclude with a comprehensive experime
 ntal evaluation where\nI also show that the performance of an object detec
 tor greatly depends\non how one defines performance. By changing the crite
 ria\, I show how\nthe ROC curves change considerably. I define three real-
 world problems\nand two well-matched criteria for optimizing these problem
 s to\nevaluate Beamer. The results show all layers of the system greatly\n
 boost accuracy.
LOCATION:Department of Engineering\, Lecture Room 5
END:VEVENT
END:VCALENDAR