Triple threat: U of T computer scientist combines machine learning, robotics and computer vision

Her research may change everything from the car you drive to the outfit you wear

Ever since she was an undergrad, Raquel Urtasun has been fascinated by machine learning (teaching computers how to think), computer vision (helping computers perceive videos and photos) and robotics.

“In many applications, you can’t disentangle these three things,” the University of Toronto computer scientist says.

After completing postdoctoral fellowships at MIT and UC Berkeley, Urtasun decided to specialize in all three, creating algorithms to help computers make decisions previously reserved for humans.

“In general, you train the computer by giving it an example and then giving it the answer to the task you want it to solve,” says Urtasun. “You show it an image of a car and then tell the computer it’s a car. You show it a dress someone thinks is fashionable and then tell the computer it’s considered fashionable.”

Programming the way to driverless cars

Urtasun is especially excited about how machine learning and computer vision could help make self-driving cars an affordable reality by programming cheap sensors to be smarter and more robust.

If perfected and accessible to consumers, autonomous vehicles could prevent accidents, reduce traffic congestion and since people won’t need to actually drive, they can spend their car time in other productive ways.

Currently, the laser-based (LIDAR) sensor used in most of today’s prototypes for self-driving cars costs approximately $80,000 US, with the perception algorithms used to guide the car relying on pre-set-up manually-annotated maps.

“Our algorithms reconstruct the environment in 3D using one or two cameras, which make for much less expensive sensors,” Urtasun explains. “We can infer the intentions of the car, the cars around it, and note other things on the road like cyclists.”

Urtasun’s work on self-driving cars has been recognized with a series of accolades and honours. She has been named the Canada Research Chair in Machine Learning and Computer Vision, received two Google faculty awards, a Connaught award, an Early Career award as well as a best paper award at the Computer Vision and Pattern Recognition conference, the premier gathering of scholars in the field.

An algorithm with style

Another example of Urtasun’s high-profile work involves using such programming for an application that may be a little less pressing for some, but more so for others.

Along with U of T computer scientist Sanja Fidler, Urtasun has created an algorithm that analyzes a person’s photograph to determine if their outfit is stylish, giving the user a rating out of a possible 10 for how fashionable they look.

The algorithm also suggests ways to improve their look and judges the subject’s overall appeal. “It will tell you things like ‘your ensemble will look better if you change your shoes to boots’,” says Urtasun.

“People care about their look, particularly when putting their picture online,” says Urtasun. “Our program is like having access to a personal assistant who can help you put your best foot forward.”

To create the algorithm, Urtasun, Fidler and their team spent a year analyzing more than 144,000 posts from the popular fashion site Chictopia and training the computer to mimic the types of choices and opinions that people were posting on the site.

With more than 250,000 users and millions of posts, Chictopia lets people upload photos of themselves, tagging the type of clothing they’re wearing (chic or casual, for example) and their location so other users can comment on and rate their outfit and overall look.

“We were really surprised by the attention our style algorithm got,” says Urtasun, who was featured along with Fidler in news, technology, science and fashion publications ranging from the UK’s Daily Mail and the Huffington Post, to New Scientist and Wired, to Glamour, Elle, Harper’s Bazaar, and Cosmopolitan.

Computer science for the masses

“We thought this was something people could use and we realized that the deep learning work we were already doing would be very useful in this application,” says Urtasun.

“What surprised me the most is how much people are interested in the idea that this is even possible, she says.

Urtasun and Fidler are looking to see if there is a business case to be made for using their work to help online merchandisers sell their products or to enable fashion experts to analyze trends as a next step for their creation in the coming months.

The Bulletin Brief logo

Subscribe to The Bulletin Brief

Faculty of Arts & Science