Photo of myself

Tom Lee

Department of Computer Science
University of Toronto

I completed my PhD at the Department of Computer Science with advisors Sven Dickinson and Sanja Fidler. I have a background in math and computer science, and I hold a M.Sc. degree from the University of Toronto and a B.Math(CS) degree from the University of Waterloo.

My PhD research was in computer vision and machine learning, specifically in image segmentation, grouping, object proposals, object detection, and structured learning.

PhD Research

Learning to generate object proposals: Object categorization in images begins with locating potential objects of interest. Bottom-up grouping cues apply to all objects and offer more discrimination than brute force testing of all possible locations. While low-level cues like color similarity are effective in many cases, mid-level cues like closure and symmetry offer greater scope. Challenges lie in developing machine learning models and algorithms that achieve high recall while preserving as much precision as possible.

Cue combination: Grouping cues are most effectively used when combined: The input image (1) is oversegmented into superpixels (2) to be grouped using perceptual cues. Using appearance alone excludes the head (3). Accounting for contour closure correctly encloses the head but strays along the fence (4). Using symmetry without closure separates the horse from the fence but leaves out the head (5). Symmetry together with closure correctly segments the entire horse (6).

Symmetric Part Detection: Typical approaches to medial symmetry assume that background is segmented away from foreground, and thus cannot be directly applied to contemporary image datasets. In this project, we reintroduce the medial axis transform (MAT) with figure-ground segmentation capability. We derive a sequence optimization problem whose dynamic programming solution allows us to detect symmetric parts in cluttered images.

Conference Publications

Learning to Combine Mid-level Cues for Object Proposal Generation,
T. Lee, S. Fidler, and S. Dickinson,
Proceedings, International Conference on Computer Vision (ICCV), Santiago, 2015.
[paper, poster]

Multi-cue Mid-level Grouping,
T. Lee, S. Fidler, and S. Dickinson,
Proceedings, Asian Conference on Computer Vision (ACCV), Singapore, 2014.
[paper, poster]

Detecting Curved Symmetry using a Deformable Disc Model,
T. Lee, S. Fidler, and S. Dickinson,
Proceedings, International Conference on Computer Vision (ICCV), Sydney, 2013.
[paper, poster]

Learning Categorical Shape from Captioned Images,
T. Lee, S. Fidler, A. Levinshtein, and S. Dickinson,
Proceedings, Canadian Conference on Computer and Robot Vision (CRV), Toronto, 2012.
[oral presentation, pdf]

Journal Publications

A Framework for Symmetric Part Detection in Cluttered Scenes,
T. Lee, S. Fidler, A. Levinshtein, C. Sminchisescu, and S. Dickinson,
Symmetry, vol. 7, issue 3, 2015.
[article, open access]


Mid-Level Cues for Bottom-Up Grouping,
T. Lee,
Doctoral thesis, 2015.

Learning Object Category Shape from Captioned Images,
T. Lee,
Masters thesis, 2011.

Peer reviewer

ECCV, Amsterdam, 2016
CVPR, Las Vegas, 2016

Teaching Assistant

Introduction to image understanding (CSC420), Fall 2014.
Algorithm design, analysis, and complexity (CSC373), Fall 2014.
Data structures & analysis (CSC263), Spring 2013, Spring 2014.
Mathematical expression and reasoning (CSC165), Fall 2011, Fall 2012.
Algorithms & data structures (CSC190), Winter 2010, Winter 2011.

Valid XHTML 1.0 Strict Valid CSS