In 1954, Fred Attneave showed a line drawing of his cat. He argued that object
shape information concentrates around high curvature locations or corners,
hence efficiently captured in a piecewise line representation.
|Attneave's Cat (1954)
A line representation is particularly attractive for articulated objects, such
as people, where individual limbs can be roughly modeled as parallel lines. In
comparison to using points, a line representation is more compact and is
better at capturing the configuration of parts.
Of course, lines are harder to work with than points, as they are more abstract
and are not well-defined physically. To use lines for recognition, we need to
answer two questions:
- How to construct a line representation?
We construct lines from bottom-up using edge detection and
constrained Delaunay triangulation, as in our ICCV '05 paper on contour grouping.
- How to match lines to lines?
A line by itself is non-discriminative; shape exists only in the
way lines are posed w.r.t. one another. Imagine matching a template (a set of
lines) to a test image (also a set of lines). If a pair of lines in the
template are adjacent (or parallel), then they should match to a pair of lines
in the image that are adjacent (or parallel). We define shape using such
relative geometric configurations between all pairs of lines. The matching can
be solved as quadratic programming.
One challenge in line matching is that there is always a certain amount of arbitrariness how we break up contours into line segments. We accommodate this in our framework by allowing "fractional" matching between lines.
|10 line aspects we learn from a skating
video sequence. A line representation is more efficient than a point-based
representation, hence covering a large variety of poses with a small set of
Sample results using these templates for single image detection: