非表示:
キーワード:
-
要旨:
Human body motion presented as point-light stimuli can be readily recognized. Psychophysical experiments show that these impoverished stimuli are sufficient for the discrimination between different actions and even for the extraction of the gender and other details of the moving actor. Additionally, a few studies indicate the capability of humans to learn to discriminate different styles of natural movements (e.g. gaits or sports movements).
However, it remains unknown whether this learning is based on innate templates
for biological movement patterns, or if humans can learn new representations of arbitrary complex movements. We address this question by investigating whether subjects can learn artificial biological movement stimuli.
Methods: The stimuli used in this study were generated by linear combination of prototypical
trajectories in space-time using spatio-temporal morphable models (Giese &
Poggio, 1999). We created two different classes of stimuli: (A) Stimuli derived by linear
combination of dissimilar natural movements (e.g. walking, kicking and dancing). (B)
Stimuli generated by animation of an artificial skeleton model that is highly dissimilar
from naturally occurring body structures. The joint angle trajectories of the skeleton
were given by linear combinations of synthetic sinusoidal trajectories. Their amplitudes
and frequencies were approximately matched with the joint trajectories of human actors
during natural movements. For both classes, several stimuli were created which served
as prototypes for the morphing procedure.
Subjects had to discriminate within one class between pairs of these stimuli that were
defined by linear combinations with slightly dissimilar weights of the prototypes. The
trajectories were presented as normal point light walkers (PLW), and as point light
walker with position jitter (PLWJ). The PLWJ were generated by adding random displacements
of the dots along the skeleton of the walker in each frame (similar to Beintema
& Lappe, 2002). Each subject took part in three test blocks that were intersected by two
blocks of training. Feedback was provided only during training.
Results: Subjects trained with stimuli derived from natural movements (class A) learned
the discrimination between novel patterns very quickly, regardless of the stimulus type
(PLW/PLWJ). If the training stimuli were rotated 90 deg in the image plane against the
test stimuli, we observed transfer of the learned representation only for the normal PLW
but not for the PLWJ stimuli.
The completely artificial stimuli (class B) were only presented as PLWJ. Subjects were
able to learn these stimuli equally fast as the natural stimuli (class A). In addition, we
found the same view dependence as for the natural stimuli.
Conclusions: (1) New templates for movement recognition can be learned very quickly.
(2) Learning affects at least two different levels of representation (local and holistic). (3)
The learned holistic representations seem to be view-dependent. (4) There seem to be no
significant differences between the learning process for stimuli derived from artificial
and natural movements.
