Keywords: morse smale complex, critical contours, shape representation
TL;DR: Goal: define topological features to make shape representation and recognition more flexible and human-like
Abstract: Computer vision is outstanding on fixed datasets, but lacks the
effortless generality and flexibility of human perception.
As a step toward understanding this flexibility, we
show that there are features -- critical contours -- readily detectable in images that have
counterparts -- extremal contours -- on 3D shapes. Both are
part of a topological structure, Morse-Smale complex, and together they provide an invariant
linking image to shape structure. An open neighborhood around these contours
is related to human perception of shape features, such as
bumps, which suggests that training for these features could provide a
generalization benefit in algorithm development.
1 Reply
Loading