ABSTRACT:

Image segmentation in computer vision has evolved such that it is routinely treated as an end task.  For example, for autonomous driving, we are interested in segmenting a road scene into {\it cars, bikes, motorcycles, persons, trees, lamp-posts, traffic signs, curbs}, etc.  To differentiate a person in different contexts, we label {\it a person on a bike} a {\it bike-rider}, {\it a person on a curb} a {\it pedestrian}, {\it a person on a horse} a {\it horse-rider}.  To understand the intent and action of a person, we want to segment a person into {\it head, torso, arms, legs}.  Segment-Anything-Model (SAM) takes supervised segmentation to a large scale, giving a false impression that segmentation is now solved.

My view is that segmentation underlies the generalization capability of visual intelligence and supervised segmentation is simply the wrong approach.  Segmentation should be treated not as an end-goal itself, but as an internal mid-level representation that serves visual recognition.   I will present our recent works in this direction, including unsupervised learning of objectness and visual context, unsupervised discovery of visual semantic hierarchies and part-whole hierarchies.

BIO:

Stella Yu received her Ph.D. from Carnegie Mellon University, where she studied robotics at the Robotics Institute and vision science at the Center for the Neural Basis of Cognition.  Before she joined the University of Michigan faculty in Fall 2022, she has been the Director of Vision Group at the International Computer Science Institute, a Senior Fellow at the Berkeley Institute for Data Science, and on the faculty of Computer Science, Vision Science, Cognitive and Brain Sciences at UC Berkeley.  Dr. Yu is interested not only in understanding visual perception from multiple perspectives, but also in using computer vision and machine learning to automate and exceed human expertise in practical applications.

Hosted by Professor Bolei Zhou