Shape Segmentation with Randomized Isoline Cuts

MPhil Thesis Defence


Title: "Shape Segmentation with Randomized Isoline Cuts"

by

Mr. Menglin CHEN


Abstract:

Segmentation of 3D surface meshes into meaningful parts is a fundamental
problem in computer graphics. Geometry applications such as skeleton
extraction, texture mapping, morphine, shape retrieval, matching, and
modeling, often involve segmentation as an initial step. However, automatic
decomposition of a mesh into meaningful components that match human
intuition is a hard problem, as it is difficult to define a measure that
captures the semantic information of a given shape.

According to the state-of-the-art benchmark on automatic mesh
segmentation algorithms, a recent randomized cut technique produces
segmentation results that are more consistent with human-generated
segmentations than other recent methods. Their method iteratively splits
the object by randomly finding a large set of possible cuts and
selecting the one with the best boundary. However, since a large set
of possible cuts are considered for each split, there is a tradeoff
between quality and computation cost, requiring several minutes to
process a single model.

In this thesis, we propose an automatic segmentation method based on
randomized isolines on 3D mesh surfaces. The isolines are
iso-contours sampled from the segmentation fields associated with
critical points located at prominent extremities of the model.  The
segmentation fields are obtained by solving constrained Laplacian
systems using a novel weighting scheme which we call the inverse
Gaussian weighting scheme. The resulting fields have the desirable
property of exhibiting large variation at concave regions while
being mostly constant on convex and flat regions. Consequently, the
uniformly sampled isolines are much denser at concave regions,
coinciding with where humans are likely to segment the model. Our
method is fast and generate segmentation results that better match
human-generated segmentations than most existing methods.


Date:       		Wednesday, 10 March 2010

Time:               	1:00pm - 3:00pm

Venue:               	Room 3588
 	                lifts 27/28

Committee Members:   	Prof. Chiew-Lan Tai (Supervisor)
                     	Dr. Hua-Min Qu
                     	Dr. Pedro Sander


**** ALL are Welcome ****