Prof. Trevor Darrell, trevor@eecs.berkeley.edu

Spring 2013

2 Units

LOCATION AND TIME CHANGE: Newton room, Room 730 Sutardja Dai Hall, Friday 10-12am

This course historically covers computer vision techniques for object and category recognition, as well as recognition of human activity from video streams.  Recognition of individual objects or activities (the coffee cup on your desk, a particular chair in your office, a video of you riding your bike) or generic categories (any cup, chair, or cycling event) is an essential capability for a variety of robotics and multimedia applications.  The advent of standardized datasets and evaluation regimes has spurred considerable innovation in this arena, with performance on benchmark evaluations increasing dramatically in recent years.  This course reviews methods that have achieved success on such datasets, and may also consider the techniques needed for real-time interactive application on robots or mobile devices, e.g. domestic service robots or mobile phones that can retrieve information about objects in the environment based on visual observation.  This class is based exclusively on readings from the recent literature, including those appearing at the CVPR, ICCV, ECCV, ICML, and NIPS conferences.

This year's version of the course will focus on three themes, covering eight papers in depth (and exploring implementations of each) in each of the following themes:

  • fine grained recognition
  • layered early vision representations (including, but not limited to, "deep" and convolutional models)
  • domain adaptation

Please click here for the current syllabus.

The format of the course this year has been expanded to include both discussion of papers as well as implementation and experimentation of covered methods.  Each week we will cover two papers on one of the above topics. Students will be assigned (possibly as a small team) to cover one or two papers during the term, and will be expected to 1) present the method and results of that paper, 2) obtain an implementation of methods they cover, and 3) prepare small example "teaser" experiments for the other students to try that week.  As part of the course a common evaluation testbed will be constructed within which these methods can be compared.  All students will be expected to try the teaser experiments each week.  There will be no problem sets or exams in this course.  Grades will be assigned based on the quality of in class presentation, implementation of associated methods, participation in discussions, and completion of the teaser problems.

Prerequisite: Active research effort on related topic.  Permission of instructor.