Graphics Reference
In-Depth Information
time [t]
I t
I t
RBF
max local
max local
Moti o n local
RBF
max local
max local
I t-1 I t
Motion global
RBF
max local
max local
S1
C1
S2
C2
Figure 7. Visual feature extraction. Motion and form features are processed along two separate
pathways, one form- and one motion pathway. The initial motion representation is further
subdivided to build separate representations of global and local facial motion. All three
streams, or channels, are further processed in parallel by hierarchical stages of alternating
S- and C-fi ltering stepsand fi nally combined into a single feature vector which serves as
input to the successive classifi cation stage.
(Color image of this fi gure appears in the color plate section at the end of the topic.)
Figure 8 demonstrates the capability of the approach to analyze
differential features in non-verbal communication represented in
segregated channels of visual motion information. In the case shown,
the subject moves the head to point out disagreement or even disgust.
This expressive communicative feature is encoded in the global
affine flow pattern showing head motion to the right (color coded in
accordance to the color map in Figure 8, right). The local motion activity
overall depicts a brief moment in which the person opens her eyes
(upward motion of eye lids (color code)) and also the chin region
moves left-downwards while closing the mouth. Both motion features
are now available to feed forward into the emotion classifier network
for analyzing the motion related non-verbal communication behavior.
Notice that in the residual flow pattern overall motion is reduced and
solely local motion that is caused by facial expression remains (in the
shown example caused by eye-, mouth- and cheek-movement).
 
Search WWH ::




Custom Search