Course Overview

Visual data is exploding! Billions of cell phones equipped with cameras exist today, with billions of consumer photos being taken each year world-wide. Hundreds of millions of photos are taken per year in NYC alone. Hundreds of new video-hours are uploaded on YouTube per minute. And we are not just taking pictures. Pictures are being taken all around us, including diagnostic medical images such as an ultrasound, images from traffic cameras on our drive to work, or imagery related to video safety and security systems for our homes and buildings. In the era of ‘big data’, major opportunities arise for systems that are capable of automatically analyzing, searching, and classifying content from images and videos. The goals of this course will be to understand the state-of-the-art computer vision approaches in the field of visual recognition and search, to actively analyze their strengths and weaknesses, and to identify interesting open questions and possible directions for future research. The course will consist of lectures given by the instructors and paper presentations by students. We will also ask students to implement cutting-edge techniques for visual recognition and search, as part of their term projects.

We will cover the following topics:

• Low-level feature descriptors, feature coding, and pooling
• Part-based and attribute-based representations
• Hierarchical models (e.g., deep convolutional neural networks)
• Large-Scale image classification and retrieval
• Smart indexing via learning-based hashing
• Active learning for efficient image retrieval and classification
• Applications

See the course schedule for details.

Prerequisites

Background in Computer Vision or Digital Image Processing is required. Programming skills in Matlab or C/C++ are also required. This is an advanced computer vision course for graduate students only.

Grading

Class Participation (10%) Paper Presentations (20%) Paper Reviews (20%) Term Project (50%)

Projects

Projects may be done in teams of two or three, depending on the total number of students enrolled in the course. Each team will have to write a paper (4–8 pages) as a final technical report. The contribution of each team member will have to be clearly specified in the project presentation. Instructors will provide links for benchmark datasets as well as for publicly available state-of-the-art implementations (source code), which can serve as basis for projects. Check the project page for more information.