Accuracy of 3D Scanning Technologies in a Face Scanning Scenario

1
Accuracy of 3D Scanning Technologies in a Face
Scanning Scenario
Chris Boehnen and Patrick Flynn University of
Notre Dame
2
Overview

• Scanner Overview
• Defining Scanner Performance
• Experiment Method
• Results
• Conclusions
• Future Work

3
3Q From 3DMD

• Designed for faces
• Uses four cameras in two camera pods with a
  flash pattern to capture the 3D data in 2
  milliseconds
• Two camera pod system allows for coverage from
  ear to ear, but limited focal range

4
FR1

• Vendor left anonymous because evaluated scanner
  does not represent their most current technology
• Uses a projected infrared pattern and a single
  camera to capture the 3D in real time
• Designed for faces and biometrics
• Fixed focal range
• Infrared pattern makes it very robust to
  lighting variation

IRCamera
IR Projector
5
FR2

• Vendor left anonymous because evaluated scanner
  does not represent their most current technology
• A series of visible light patterns are projected
  over 20 seconds and 3 cameras calculate the 3D

• Has a limited focal range and is sensitive to
  lighting conditions

Camera
Visible Light Projector
6
Konica Minolta 910

• Utilizes a camera and laser stripe
• Needs subject to hold still for 2.5 seconds
  during scan
• Different zoom lenses help give it a large focal
  range
• Works well in different lighting conditions

7
Polhemus Fastrak

• Uses a magnetic field to track the location of
  the wand (fixed laser stripe and camera) which is
  manually moved over the object
• Person must stay still during scanning which
  takes 30s depending on quality and operator

8
Defining Scanner Performance

• Level of detail
• Repeatability
• Accuracy to scanned object
• Percentage of object captured
• Texture image
• Impact of environment

9
Accuracies We Explore Here

• Absolute Accuracy - the accuracy of the scanner
  with respect to a physical reference object
• Repeatability Accuracy - the degree of
  consistency between scans
• Pose Variation Accuracy As the orientation of
  the scanner to the object changes how does the
  data change

10
Experimental Method

• Use synthetic face masks of known ground truth as
  scanned objects
• Allow each scanner to perform the scans under
  its own ideal conditions
• Compare resulting scans to each other and known
  ground truth

11
Synthetic Face Masks

• An initial scan of a face from the Minolta is
  cleaned to produce the ground truth for the mask
• 10 faces, five female and five male from
  different ethnicities are used for the initial
  scan
• A mold for the mask is produced using a Thermojet
  rapid prototyping machine from 3D Systems and a
  durable 2 part polymer is poured in
• The ground truth was verified with a tactile
  scanner and found to be 0.018mm

12
Data Collection Overview

• Capture a series of scans on each scanner under
  their ideal conditions
• Run Raindrop Geomagics cleaning algorithm on the
  scans
• Perform an initial alignment to the scans and the
  ground truth

13
Data Capture
14
Comparing Scans

• Visualization Toolkits implementation of ICP is
  run to align the scans as closely as possible
• The distance from each point on one surface to
  the other is used to calculate the mean and Root
  Mean Square (RMS) error distances

15
Absolute Accuracy Comparisons
16
Repeatability Accuracy Comparisons
17
Results Overview
Error values are in mm
18
Results Pose Change
Error values are in mm
19
Outlier Detection and Removal

• We identified outliers as anything with a
  distance of over 20mm

20
Conclusions

• The most accurate scanner here in every category
  is the Minolta followed closely by the 3DMD and
  Polhemus with the anonymous vendors far behind
• In a real world setting the 3DMD may provide
  slightly more accurate data with live moving
  subjects, but has a sufficiently lower LOD that
  we ranked it second here

21
Conclusions

• The Polhemus scanner provided good data and is
  the smallest scanner here by far leaving open the
  possibility for a portable version
• FR1 provided a decent accuracy rating but the
  number of points was so low that we ranked it
  fourth here
• FR2 has a high LOD but the accuracy of those
  points is very low and is ranked last here

22
Future Work

• Addition of more scanners
• Exploring impact of color on accuracy
• More realistic representation of skin material in
  face masks

23
Questions?
24
(No Transcript)
25
Verifying the Ground Truth

• The Thermojet lays semi-rectangular dots of
  material with a diagonal of 0.11mm
• At worst parts of the model are only accurate to
  half the diagonal, but on average the model
  should be accurate to around .03mm
• Using a tactile scanner operating on a 100 micron
  grid (1.5 million points), and accounting for
  systematic scanner error, we verified the
  models to be within approximately 0.018mm

26
Graphical Depictions of Accuracy
27
Scanner Review