�ݺ�ߣ

Abstract
Humans have an innate ability to connect and understand
complex visual data. Specifically, Gunnar Johansson (1973) showed
that we can recognize activities of people only from movements of
their joints. This kind of biological motion is an intrinsic part of
how our brain decodes visual information. The purpose of this
study is to pinpoint the aspects of natural human sight that are
most important for understanding biological motion.
Background
• Biological motion is the unique style of animation that is
displayed by moving objects.
• Point Light Display is the method of recording an object or
creature moving with lights attached to their joints. The product
is an image or video of just the joints.
Methods
Point Light Display
• Ten videos were recorded with actors who had ten LED lights
attached to their joints (Figure 1) and were varied in complexity
of action and number of actors (1, 2 or 3).
• Type of actions: walk, shake hands, bow, skip, jump, high‐five,
kick
• Recordings were processed to extract LED light locations and
reduced to point sets.
• Distortion types of changeable degrees were applied to the point
sets and new videos with the distortion were generated. (Figure
2)
Challenging the Cognitive Capability to Connect the Dots:
Recognition of Biological Motion with Point Light Display
Aviel J. Stein and Toshiro Kubota
Department of Mathematical Sciences
Results
• The video survey was taken by 168 people and 1676 videos
watched.
• The Control group had a 100% full recognition of video while the
overall full recognition of video in test videos was 92.7% , 4.6%
partial recognition, and 2.6% no recognition.
• Table 2 summarizes the recognition performance.
• USD was effective, but subjects were often able to learn and
adjust.
• Perturbation on size (Pulse, Size) did not affect the results.
• Large jitter (Jitter‐30) was effective to some degree.
• Subtle actions (bowing, high five) and quick actions (kick) were
more vulnerable to disturbance.
Discussion
• The results demonstrated robustness of the human vision (92.7%
full recognition of disturbed videos).
• We expected recognition of biological motions to be sensitive to
temporal sampling, but we were proven wrong; Skip‐8 still
retained 88% full recognition.
• No local perturbation appears to be effective in disturbing our
perception.
• No systematic perturbation (USD and Reverse) appears to be
effective in disturbing our perception.
Conclusion
Our results show resilience of the human vision against various
random and systematic distortion in recognizing biological motions.
Our future research includes algorithmic treatments of the grouping
and recognition processes, and empirical study of recognition from
biological motions beyond actions, such as gender, posture, and
animals.
References
• Johansson, G. (1973). Visual perception of biological motion,
biological motion and a model for its analysis. Perception &
psychophysics, 14, 201‐211.
TYPE Control Jitter‐15 Skip‐4 Rem‐25 Rev‐3 Pulse‐24 Size‐15
%Error 0 6.67 7 2.85 1.25 0 0
TYPE USD Jitter‐30 Skip‐8 Rem‐50 Rev‐6 Pulse‐48 Size‐30
%Error 20.67 16.4 12.24 2.85 15.62 1.11 0.91
Table 2. Combined percent for no and partial recognition.
Figure 1. The image on the right shows an actor who moves from left to
right and on the left is the actor who moves from right to left.
Table 1. Descriptions of the types and degrees of disturbance.
TYPE Description
USD Frames��flipped��up‐side‐down
Jitter‐k Each��dot��is��perturbed (on��both��x��and��y)��by��random��
noise��uniformly��distributed��in��[‐k,��k].
Skip‐k
Every��k‐th frame��followed��by��k‐1��blank��frames��are��
shown
Remove‐k k��percentage��of��the��dots��are��removed
Reverse‐k Each set��of��k��frames��are��reversed��in��order
Pulse‐k
The��diameter��of��dots��grow/shrink��periodically��
between��[1��k].
Size‐k
The��diameter��of��each dot��is��set��randomly��to��a��value��in��
[1��k].
Figure 3. Illustrations of disturbances
1 2 3 4 5
1 1BBlank 1BBlank 1BBlank 5
4 3 2 1 8
1 2 3 4 5
Control
Skip 4
Reverse 4
Pulse
USDControl Jitter Remove Size
Distortion
• Seven types of distortion: upside‐down (USD), Jitter, Skip,
Remove, Reverse, Pulse, and Size (See Table 1 and Figure 3).
• Two levels of distortion (except USD): Jitter‐15, Jitter‐30, Skip‐4,
Skip‐8, Remove‐25, Remove‐50, Reverse‐3, Reverse‐6, Pulse‐24,
Pulse‐48, Size‐15, Size‐30.
Figure 2. Examples of generated video frames.
Acknowledgement
This work was supported by NSF grants CCF‐1117439 and CCF‐
1421734.

�ݺ�ߣ

Stein_Poster_1.7

Recommended

More Related Content

Featured (20)

Stein_Poster_1.7