Sensory motor coordination in gaze control

1
Sensory motor coordination in gaze control
2
Introduction
3
Introduction
Fixating the eye
4
Gaze Control

• Why are we interested in gaze control?

5
Gaze Control
6
Gaze Control
7
Gaze Control

• Non-situated models (open loop)
• pre-determine fixation locations
• Emphasis on high-level tasks
• Situated models (closed loop)
• base gaze control on domain knowledge
• learn gaze control
• Emphasis on low-level tasks
• Important concept Sensory-motor coordination

8
Gaze Control

• Can sensory-motor coordination contribute to
  performance of situated models on high-level
  tasks?

9
Non-situated Model
10
Non-situated Model
11
Situated Model
12
Situated Model
13
Experiment

• Compare non-situated and situated model on gender
  recognition
• 278 color images of human subjects ? 600 x 800,
  gray-scale
• 15 evolutionary runs per type of model on
  training set
• Each model can perform 5 fixations per image
• Performance
• Average ratio correct classifications per image
  (test set)

14
Results
Non-situated model
Performance Average ratio correct
classifications per image (on test set)
Situated model
15
Results
Categorisation ratio class-conditional
performance at a location
Male images
Female images
16
Results
Class-dependent gaze behaviour
Categorisation ratios of male images in the
training set.
Categorisation ratios of female images in the
training set.
17
Results
Sensory-motor coordination enhances performance
over multiple time steps
18
Discussion

• Enhancing computer vision
• Understanding human gaze control

19
Conclusions

• Sensory-motor coordination contributes to
  performance of a situated model on the high-level
  task of artificial gaze control for
  gender-recognition in natural images.
• 2. Sensory-motor coordination optimises
  categorisation performance by establishing useful
  dependencies between multiple actions and
  observations

20
Future work

• Explaining sensory-motor coordination within a
  Bayesian framework
• Employing sensory-motor coordination with various
  classifiers