Graphics Reference
In-Depth Information
the components still have defects on their border, for which we presented a new blending
technique that retains the full correspondence between the face instance and the morphable
model. These newly generated face instances can be added to the example set from which a
new and enhanced morphable model can be built. We tested our fully automatic bootstrapping
algorithm on a set of 16 new face scans, to show that the new morphable face model has
indeed an enhanced expressiveness. By adding data to an initial morphable face model, we can
fit the model to face scans of a larger population and reduce the confusion among identities.
Therefore, the building of a strong morphable face model is essential for the difficult task of 3D
face recognition. However, adding data to a morphable model increases the computation costs
for both model fitting and face comparison. Therefore, we developed an automatic redundancy
check, which discriminates between new face data to add to the model and redundant face
data to reject. On the basis of the initial 16 scans, we selected a residual error threshold for the
automatic redundancy check. Then, we applied our bootstrapping algorithm to 277 UND scans.
Our bootstrapping algorithm successfully established full correspondence between these scans
and the initial morphable model, and selected 35 persons that were new to the model.
With our new bootstrapping algorithm, we are able to successfully update an initial face
model, which we use to produce more accurate fits to new scan data. The algorithm is fully
automatic, reuses initial face statistics, checks for redundancy, and retains the full correspon-
dence even in case of noisy scan data with holes. The algorithm successfully enhances the
neutral morphable face model with new (close to) neutral face scans. It should also apply to,
for instance, a surprised face model and surprised scans, but not to a neutral face model and
surprised scans. To establish correspondences among face scans with different expressions,
new automatic algorithms are required.
Face matching using facial contours shows higher recognition rates on the basis the multiple
component fits than for the single component fits. This means that the obtained 3D geometry
after fitting multiple components has a higher accuracy. Our face modeling approach in
combination with the three selected contour curves achieves 98% RR on the UND set. The
obtained model coefficient that were used to produce the accurate face instances, turned out
to have the highest performance. With the use of four components, we achieve 100% correct
identification for 876 queries in the UND face set, 98% for 244 queries in the GAVAB face
set, and 98% for 700 queries in the BU-3DFE face set. These high recognition rates based
on normalized coefficient vectors proves that our model-fitting method successfully fits the
morphable face model consistently to scan data with varying poses and expressions.
Exercises
1.
The pose normalization based on fitting a coarse nose template fails when a scan is made
from the side (such that half the nose is occluded), or when the scanning range is not right
(such that the nose is not scanned). (a) Think of a way to detect such cases. (b) Develop an
alternative way for pose normalization that does not suffer from such cases.
2.
When a morphable model is fitted to a frontal face scan, the resulting weight values are
not well commensurable to the weight values resulting from fitting a morphable model to
a side face scan. Invent a strategy to cope with this situation.
