This course will cover current topics in computer vision with an emphasis on motion estimation, learning methods, and probabilistic models. Readings will be from recent research papers. Computational techniques such as Expectation-Maximization, Hidden Markov Models, and Belief Propagation will be introduced. Applications to tracking, recognition, image enhancement, and human motion analysis will be considered. Prerequisites: basic probability, linear algebra, and calculus.
Each year this course focuses on a different area of computer vision with the focus on a real application. There has been a renewed interest in object recognition in the vision community. Despite recent advances, there are no current vision systems that could recognize 100 household objects in clutter, varied lighting, with occlusion, etc. Moreover, many recent methods have focused learning to recognize objects based on the statistics of some image features. These methods typically do not model the 3D shape of the object.
This class is being offered in collaboration with Willow Garage. Willow Garage is building a mobile household robot that will be able to do many things, among which are tasks like clearing the table. To do so the robot must not just detect objects but recognize their 3D pose so that it can grasp them. Consequently this class will focus on the problem of object recognition in this context.
The class will inolve reading recent papers and implementing an object recogntion system. CS296-4 is always about learning by doing. The class is structured as a big project and class time is a lot like a big research group meeting. Past classes have been fun and a lot of work.
We will cover:
- machine learning approaches to object categorization
- image features, shape contexts
- bottom-up object detection
- top-down 3D model fitting
- illumination and shading
Everyone in the class will get a couple of objects of their "own". The objects include glasses, cups, plates, cuttlery, etc from IKEA. Along with your object you will receive a 3D mesh model of the object obtained by a laser range scanner. Your goal will be to recognize your objects (and not others) in a range of increasingly difficult conditions.
Competition and prizes: at the end of the course we will have a set of "challenge" problems for everyone's system. Prizes for the best performing systems will be provided by Willow Garage.
To see the preparation of the object database click here.
A computer vision course, e.g. CS143 (Introduction to Computer Vision) or an equivalent course.I'll assume good familiarity with linear algebra, calculus, probability, statistics, (e.g. CS155, AM0040, AM165, AM169, or AM264).
Matlab programming experience.
Michael Black
Office: CIT 521
Email: black<at>cs<dot>brown<dot>edu
Office Hours: Mondays 2:00-3:00 and Tuesdays 5:00-6:00
TA: Alexandru Balan
This is not a "normal" class. It will be messy and hard. There will be few lectures, no textbook, no toy assignments, etc. The goal is to solve a real problem that will make an impact. We may fail. My role is to lead us all through this process to hopefully a successful outcome.Participation (10%) – absolutely necessary. If you don’t come to meetings you won’t get anything out of this course.
Paper presentations (20%) – We'll read a lot of papers. Everyone will present several papers and must be prepared to discuss the others; everyone must email me a paragraph describing each reading and point out at least one problem with it.
Project (70%) – This will be broken into pieces and I will give you bits and pieces of the solution. There is no one right answer or approach. Everyone will present their methods in class as they are developed.
You may use any materials you can find (e.g. code from the web) and you may talk with anyone in the class. You must do your own work (e.g. programming and experimentation) but I encourage you to talk with and learn from each other. You may also seek help from outside the class but you must ask permission before sharing the data. I expect you must uphold the highest standards of scientific conduct which includes the proper assignment of credit.
