DentaltownUK May 2017 - 59

hollywood smile

Thabo went on to say: "Our core focus
was to be extraoral since we don't want to
put our actors through the invasive procedure of intraoral scanning, and since we're
concerned with entertainment purposes,
we're interested in faithfully capturing
only those parts that will be visible on the
screen."
Disney's algorithm relies on several factors, including the shape, pose and scale
variation of each individual tooth. It
requires minimal user interaction and can
operate on a set of uncalibrated images or
even a short video clip. Images and videos
taken on both cameras and handheld
devices like an iPhone can provide a suitable basis to create the model, which was
proven to yield strong results.
Because dentists spend a lot of time
standardising and calibrating photos, Halley was especially intrigued by this aspect.
She wondered, 'How many images at a
minimum are needed, and would you need
a series of expressions to capture the gum
line?' While Thabo admits to it not being
investigated thoroughly (which is fair considering the project is still being worked
on), the team has experimented with
reconstructing teeth from even a single
photograph and managed to get plausible
results. This is still incredibly impressive.

"Times and conditions change so rapidly
that we must keep our aim constantly
focused on the future." - Walt Disney
Although it doesn't always have a 100
percent success rate, the project is currently
based on statistics and hypothesis, which
we will explain further. With more tweaks
and research, this can only improve.
To read the full research paper on the
algorithm and processes at Disney Research
Labs, visit https://goo.gl/lWLIAV. Here's a
quick summary:
To begin, the team collected a database
from 86 high-resolution plaster casts of
oral 3-D scans of a variety of upper and
lower teeth. They also created individual
tooth template meshes, categorised into
molars, premolars, canines and incisors.
They modeled these meshes all the way to
the root of the tooth to give any reconstructed model a solid base.
This is then combined with the data
gained from a provided image of the teeth
to create a reconstructed model. The image
is used to detect the shape and boundaries
of the teeth. The outline of the teeth is
taken and combined with the template
mesh and the database to calculate where
the teeth would naturally and most likely
be positioned.

This also allows the algorithm to identify
the positioning of teeth that may not be
included in the image (usually molars). The
use of sparse image data can be justified
and worked around, should it be required,
to create an accurate reconstructed model.
Thabo summarises: "Our system consists, broadly speaking, of two main steps.
First, it estimates the teeth boundaries
from the images using a machine learningbased feature classifier. Second, it optimises
for the teeth shape and pose that best fit the
extracted features while being faithful to
the statistical distribution of the prior
model.
"The algorithm will hypothesise the
most plausible teeth where they are missing

dentaltownuk.com \\ MAY 2017

59
https://www.goo.gl/lWLIAV https://s3-us-west-1.amazonaws.com/disneyresearch/wp-content/uploads/20161123192449/Model-Based-Teeth-Reconstruction-Paper.pdf http://www.dentaltownuk.com
Table of Contents for the Digital Edition of DentaltownUK May 2017
