3D Morphable Models for Face Surface Analysis and Recognition - 3D Face Modeling, Analysis and Recognition - page 136

To elaborate on the performance of our bootstrapping algorithm, we applied it to the data
set of 16 different face scans and to the subset of 277 UND scans. The small set is used to
evaluate the model-fitting and correspondence estimation algorithm. The UND set is used to
test the redundancy estimation.

Correspondence Estimation

To evaluate the correspondence estimation, we compare the residual errors of fitted face
instances. First, we fit the initial morphable face model as a single component to the segmented
face data S single . Second, we fit the initial model to the segmented face data using the four
components and blend their borders. Third, we add these 16 face instances S mult to the example
set and recompute the PCA model, keeping the m

99 principal components of the face.
Finally, we fit the enhanced morphable face model to the same segmented face data using a
single component S single .
In Table 4.3, we report the residual errors of the fitted face instances. We use these errors
for quantitative evaluation of the produced fits. For all face scans, S mult has a lower residual
error than S single , which means that a higher fitting accuracy is achieved with the use of
multiple components in combination with the improved correspondence estimation. After the
model was enhanced with the 16 face instances S mult , the model was again fitted as a single
component to the 16 face scans. Now, all residual errors are lower for S single than they were for
S single , which means that our bootstrapping algorithm successfully enhanced the morphable
face model. For one face scan, the face instance S single is even more accurate than S mult . This is

=

Table 4.3 RMS errors (mm) of output models to input scans. The model fits with the smallest and
largest difference in residual errors (bold) are show in Figure 4.10

S single S single − S mult

Data set S single S mult S single − S mult

GAVAB 1.36 1.24 1.27 0.13 0.04

GAVAB 1.34 1.16 1.19 0.18 0.03

GAVAB 2.04 1.59 1.56 0.45 − 0.03

GAVAB 1.32 1.16 1.19 0.17 0.03

BU-3DFE 1.18 1.01 1.02 0.18 0.02

BU-3DFE 1.28 1.12 1.15 0.16 0.03

BU-3DFE 1.06 0.96 0.96 0.10 0.00

BU-3DFE 1.61 1.40 1.49 0.21 0.09

local 0.70 0.55 0.58 0.15 0.03

local 0.89 0.68 0.69 0.21 0.01

local 0.79 0.62 0.67 0.17 0.05

local 0.69 0.55 0.58 0.14 0.04

CAESAR 1.86 1.77 1.80 0.09 0.03

CAESAR 1.88 1.77 1.78 0.11 0.01

CAESAR 1.81 1.74 1.77 0.07 0.03

CAESAR 1.80 1.75 1.78 0.05 0.03