CS 764 Seminar in Computer Vision

1
CS 764 Seminar in Computer Vision

• Attention in visual tasks

2
Two different tasks

• Perception of motionExperiments indicate systems
  similar to the ones used in computer vision
• Object representations and visual memoryResults
  indicate systems different fromanything we
  implement

3
Attention and motion

• Todd Horowitz and Anne TreismanSpatial Vision
  Vol. 8, No. 2, 1994
• Apparent x Real motion
• Short range x Long range system

4
Short range x Long range system

• Short rangeSmall displacement Small time
  interval between stimuliConstant contrast
• Long rangeEverything else

5
Theoretical framework

• Short range motionArray of simple motion
  detectors operating in parallel. (like optical
  flow computation)
• Long range motionDepends on linking stimuli
  separated in space and time. Requires
  attention(like shape matching)

6
Experiment 1

• Detection of long and short range motion with
  unicontrast and bicontrast stimuli
• Measure response time as a function of number of
  distractors.Parallel search gt constant
  timeSerial search gt linear time

7
Experiment 2 - Motion Aftereffect

• The effects of adaptation are reversed between
  short and long range motion.
• Short range system becomes less sensitive to
  adapted motion
• Long range system benefits from practice stimuli

8
Summary

• Two different systems for motion perception We
  have decent models for both systems
• Short range motion seems much like optical flow
  computation
• Long range motion seems much like shape matching

9
Object Tokens, Attention, and Visual Memory

• Anne Treisman and Brett DeSchepperAttention and
  Performance XVI, 1996
• Probe bottom-up registration of unattended
  objects
• Specially novel objects, for which new
  representations may be constructed

10
Introduction

• On recognition tasks, we cant remember novel
  unattended shapes
• But indirect measures (i.e. priming) reveal
  memory beyond conscious level
• How much processing and learning happens without
  attention?

11
Experiment 1

• Tests indicate unattended shapes cause negative
  primingFew subjects show facilitation
• Very long duration of effect (gt weeks)
• No explicit memory of unattended shapes

12
Experiment 2

• Are there differences between representations
  built for attended and unattended shapes?
• Maybe attended shapes are processed to a higher
  level and representations reflect this

13
Summary

• No explicit memory of unattended novel shapes
• Priming indicates implicit memory
• Implicit memory learns with one example and
  repetition is irrelevant.
• Effects last for long periods (novel shapes)
• When attended, novel shapes are processed to a
  higher level.