NonRigid transformation (0°)

geographertonguesAI and Robotics

Nov 30, 2013 (3 years and 7 months ago)

133 views

1. Introduction

The presence of familiar object motion is known to facilitate visual
object recognition
1,2
. In addition to shape cues, observers can
utilise familiar visual changes in an object’s physical properties for
identification.


Such changes typically arise in the natural environment as the
result of either a moving observer or object (Fig. 1). Here, we
show that characteristic dynamic information can sometimes
compensate for when object recognition is impaired by unreliable
shape cues e.g., 2D shape distortions by viewpoint variations
3
.

[1]
Stone, J.V., (1999), Object recognition: view
-
specificity and motion
-
specificity, Vision Research, 39, 4032
-
4044.


[2]
Vuong, Q.C. & Tarr, M.J., (in press), Structural similarity and
spatiotemporal noise effects on learning dynamic novel objects,
Perception.

[3] B
ülthoff, H.H. & Edelman, S., (1992), Psychophysical support for a
2
-
D view interpolation theory of object recognition, Proceedings of
the National Academy of Science, 89, 60
-
64.

Recognising dynamic objects across viewpoints

Lewis Chuang
1
, Quoc C. Vuong
1
, Ian M. Thornton
2
, Heinrich H. B
ülthoff


1
MPI for Biological Cybernetics, Tübingen, Germany

2
Department of Psychology, University of Wales Swansea, U.K.


MPI FOR BIOLOGICAL CYBERNETICS

4. Discussion

5. References

2. Method

Part 1: Learning


Participants learned 2 moving objects. The 2 objects changed in a

consistent fashion, either rigidly (Expt 1,3) or nonrigidly (Expt 2,4) across time.


Part 2: Two interval
-
forced choice Test


Two objects were presented in succession: one learned and one novel.

Participants had to identify the object previously learned in Part 1.



Test variables


a)

Object
-
Motion: Learned objects were presented moving in either

the learned or reverse direction (See Figure 2.1 for details).



b)

Viewpoint variation: Objects were viewed from novel viewpoints

that varied from the learned viewpoint in steps of

10
°

(
Expt 1 & 2:

0
°
, 10
°
, 20
°
, 30
°
); 20
°

(Expt 3 & 4: 0
°
, 20
°
, 40
°
, 60
°
)


Figure 2: Example of a novel object’s rigid/nonrigid transformation, across different

i) object
-
motion; ii) viewpoints.

Learned motion

Reversed motion

NonRigid transformation (0
°
)

NonRigid transformation (20
°

CW)

Rigid transformation (0
°
)

Rigid transformation (20
°

CW)

It is well
-
known that visual object recognition is dependent on:



the availability of familiar object motion [1, 2].



view
-
familiarity [3]


Both findings are replicated in our studies.


More importantly, we show that this recognition benefit of familiar

object motion remains constant across the varying viewpoints.


This is true for both rigid and nonrigid types of object motion.


Our findings highlights the relevance of familiar object
-
motion as a

cue for object recognition, that compensates for instances when

other cues e.g., shape, are less reliable.

3. Results

There is a general benefit of recognising a learned object presented in its
learned motion.

(Experiment 1: F(1,23)=13.9, p <0.05; Experiment 2: F(1,23)=8.8, p<0.05 )


Object recognition decreases with increasing variation between test and
learned viewpoint.

(Experiment 1: F(3,69)=4.4, p <0.05; Experiment 2: F(3,69)=27.0, p<0.05)


There was no interaction effect between the factors of Object
-
motion and

Viewpoint
-
variation.


Experiment 1: Rigid motion
65
70
75
80
85
90
95
100
0
10
20
30
Viewpoint variation
Mean accuracy (%)
Learned motion
Reversed motion
Experiment 2: Nonrigid motion
65
70
75
80
85
90
95
100
0
10
20
30
Viewpoint variation
Mean Accuracy (%)
Learned motion
Reversed motion
Figure 1: Real
-
life examples of rigid and nonrigid visual changes

resulting from: i) moving observer (top row) ; ii) animate object (bottom row)