2/21/2022 Aaron Seidlitz, Illinois CS
In the years since Illinois CS professor Svetlana Lazebnik earned her PhD here in 2006, she has witnessed a great deal of change to her research field of computer vision. She counts her students and colleagues as the reason why she remains so inspired in this line of work.
Written by Aaron Seidlitz, Illinois CS
There have been a couple moments in Illinois Computer Science professor Svetlana Lazebnik’s career that have allowed her the opportunity to reflect upon her own growth, as well as her niche in the computing industry.
Considering the demands of being a faculty dedicated to her craft – through countless hours in the classroom and research lab – these moments don’t present themselves very often.
Most recently, this opportunity came toward the end of January, when Lazebnik found out the University of Illinois System honored her as a University Scholar. She was one of five professors to earn the distinction for the past two years, which recognizes “faculty excellence and provides $15,000 to each scholar for each of three years to enhance his or her academic career.”
“I do feel very honored by becoming a University Scholar,” said Lazebnik, who focuses her research and academic activity on a subfield of Artificial Intelligence (AI) called computer vision. “I love this university, and I feel like it's my home. So, I do feel very grateful that Illinois recognized me in this way. I also hope I can return the favor by continuing to contribute to the high level of academic and research-based achievement at Illinois CS and throughout the university.”
In that moment, she also reflected on just how significantly her field has grown over the two decades of her research career.
And there was no better point at which to do this than two years ago, when Lazebnik delivered a series of lectures during her sabbatical.
“At that time, I spoke about the history of computer vision, in part because I wanted to make sense of it myself,” Lazebnik said. “Considering that the field has been around, roughly, since 1960, I realized that by now I have been involved in about one-third of my field’s history. But the field back then is not what it is now.
“These days I find myself thinking about that more and more, because computer vision has undergone explosive change and growth over the last 5-10 years, especially.”
The importance of this change plays a part in the way Lazebnik found a place within it.
Computer vision is the study of methods through which computers can extract meaningful information from digital images, videos, and other visual inputs.
When Lazebnik earned her PhD from Illinois CS in 2006, the field seemed more wide open to interpretation and relatively few methods existed that really worked.
But the boom for computer vision work in recent years, Lazebnik detailed, resulted in large-scale vision systems that have been deployed in the real world.
“We now have huge neural networks running on millions of images,” Lazebnik said. “As researchers, we need to produce evaluations, benchmarks, and extensive ablation studies to get a paper published. So, it's really been an incredible transformation.”
Within this period of extensive growth, Lazebnik found her success in a measured way.
Rather than pursuing highly visible and competitive benchmarks, she enjoys working on projects that hold up over time and advance the field in a valuable way.
This has led to several instances in which her work garnered a lot of citations because of the useful nature of her findings.
“I found out how much I enjoyed this method of working, really, through my PhD work,” Lazebnik said. “This was in 2006, and my work served as the very beginning of general recognition benchmarking. I developed a Spatial Pyramid Matching method that was actually a bit of a Hail Mary for me at the time, because a much more complicated project just wasn’t working out shortly before the conference deadline.
“I tried this idea instead, and, suddenly, realized that it yielded a 20% performance improvement.”
In 2016, this paper received the Longuet-Higgins “test of time” award at CVPR.
Another of her career highlights came from work done in 2015, and it resulted in a paper – entitled “Flickr30k Entities: Collecting Region-to-Phrase Correspondences for Richer Image-to-Sentence Models” – that became influential through many citations.
Together with a PhD student, Bryan Plummer (PhD '18), and fellow Illinois CS professor Julia Hockenmaier, this work became an early benchmark for what’s known as visual grounding.
“When I started collaborating with Julia, we extended a dataset she had created to go beyond whole image and whole sentence similarity,” Lazebnik said. “We wanted to understand the correspondence between pieces of sentences and pieces of images. This meant identifying which phrase from a sentence goes with which region in an image to produce better understanding.”
Lazebnik has found the most fulfilling part of her time at Illinois CS to be the interactions with colleagues and students that lead to consistently meaningful and productive research.
That personal approach to mentoring students, combined with the usefulness of her projects, is part of the reason for her University Scholar recognition.
“One of the most important motivations to be a professor, the thing that gives me a long-lasting sense of satisfaction, is working alongside our students and fellow faculty,” Lazebnik said. “Especially since computer science has become so fast-paced, the relationships with students just matter more to me. Being a mentor to students is what I think of most in my role.
“To be able to do that where I earned my own PhD, at the University of Illinois, is a wonderful experience.”