combining textureandcolourcuesinvisual segmentationmsl/talks/saarelalandy10.pdf · texture color...
TRANSCRIPT
Combining texture and colour cues in visual
segmentation
Toni P. Saarela & Michael S. Landy
Department of PsychologyCenter for Neural Science
New York University
2010-08-24
European Conference on Visual PerceptionLausanne, Switzerland
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 1 / 34
Cues for visual segmentation
Luminance
Colour
Texture
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 2 / 34
Texture and colour
Both texture and colour are robust cues that can provide informationfor segmentation under varying brightness or illumination
◮ Texture: e.g., Schofield, 2000◮ Colour: e.g., Li & Lennie, 2001
Both have been studied, but mostly in isolation
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 3 / 34
Texture and colour
Both texture and colour are robust cues that can provide informationfor segmentation under varying brightness or illumination
◮ Texture: e.g., Schofield, 2000◮ Colour: e.g., Li & Lennie, 2001
Both have been studied, but mostly in isolation
Questions
If both cues are available, how are they combined?
Independent or common neural mechanisms?
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 3 / 34
Independent or common segmentation mechanisms?
Independent
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 4 / 34
Independent or common segmentation mechanisms?
Independent Common
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 4 / 34
Segmentation task: vertical or horizontal edge?
Conditions
Texture
vs.
Orient. diff. = ∆θ
Colour diff. = 0
Stimuli
• 12×12 arrays of Gabors• Gabor spatial frequency:1 cycle/degree
• Gabors have fixedluminance contrast
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 5 / 34
Segmentation task: vertical or horizontal edge?
Conditions
Texture
vs.
Orient. diff. = ∆θ
Colour diff. = 0
Colour
vs.
Orient. diff. = 0Colour diff. = ∆c
Stimuli
• 12×12 arrays of Gabors• Gabor spatial frequency:1 cycle/degree
• Gabors have fixedluminance contrast
• Added red-greencolour contrast
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 5 / 34
Segmentation task: vertical or horizontal edge?
Conditions
Texture
vs.
Orient. diff. = ∆θ
Colour diff. = 0
Colour
vs.
Orient. diff. = 0Colour diff. = ∆c
Stimuli
• 12×12 arrays of Gabors• Gabor spatial frequency:1 cycle/degree
• Gabors have fixedluminance contrast
• Added red-greencolour contrast
• Second-order cues
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 5 / 34
Segmentation task: vertical or horizontal edge?
Conditions
Texture
vs.
Orient. diff. = ∆θ
Colour diff. = 0
Colour
vs.
Orient. diff. = 0Colour diff. = ∆c
Combined
vs.
Orient. diff. = ∆θ
Colour diff. = ∆c
Stimuli
• 12×12 arrays of Gabors• Gabor spatial frequency:1 cycle/degree
• Gabors have fixedluminance contrast
• Added red-greencolour contrast
• Second-order cues
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 5 / 34
Segmentation task: vertical or horizontal edge?
Conditions
Texture
vs.
Orient. diff. = ∆θ
Colour diff. = 0
Colour
vs.
Orient. diff. = 0Colour diff. = ∆c
Combined
vs.
Orient. diff. = ∆θ
Colour diff. = ∆c
Stimuli
• 12×12 arrays of Gabors• Gabor spatial frequency:1 cycle/degree
• Gabors have fixedluminance contrast
• Added red-greencolour contrast
• Second-order cues
Procedure
• Single-interval design• 250 ms presentation• Blocked conditions• Random stripe phase
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 5 / 34
Stimulus space: Texture condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 6 / 34
Orientationcontrast
Stimulus space: Texture condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 7 / 34
Orientationcontrast
Stimulus space: Texture condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 8 / 34
Orientationcontrast
Stimulus space: Texture condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 9 / 34
0.5
1
Orientation contrast
Pro
port
ion
corr
ect
Orientationcontrast
Stimulus space: Texture condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 10 / 34
Orientationcontrast
Stimulus space: Colour condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 11 / 34
Orientationcontrast
Colour contrast
Stimulus space: Colour condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 12 / 34
Orientationcontrast
Colour contrast
Stimulus space: Colour condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 13 / 34
Orientationcontrast
Colour contrast
Stimulus space: Colour condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 14 / 34
Orientationcontrast
Colour contrast
Stimulus space: Colour condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 15 / 34
Orientationcontrast
Colour contrast
Stimulus space: Cue-combination condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 16 / 34
Orientationcontrast
Colour contrast
Decision space: Texture condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 17 / 34
Internalresponse:Orientationcontrast
Internal response: Colour contrast
Decision space: Colour condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 18 / 34
Internalresponse:Orientationcontrast
Internal response: Colour contrast
Decision space: Cue-combination condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 19 / 34
Internalresponse:Orientationcontrast
Internal response: Colour contrast
Decision space: Sensitivity
d'colour
d' texture
d'combined
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 20 / 34
Prediction
d′
combined =
√
d′2texture + d
′2colour
Internalresponse:Orientationcontrast
Internal response: Colour contrast
Results
Texture0
0.5
1
1.5
2
2.5
d’
S1
Texture0
0.5
1
1.5
2
2.5
d’
S2
Texture0
0.5
1
1.5
2
2.5
d’
S3
Texture
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 21 / 34
Results
Texture Color0
0.5
1
1.5
2
2.5
d’
S1
Texture Color0
0.5
1
1.5
2
2.5
d’
S2
Texture Color0
0.5
1
1.5
2
2.5
d’
S3
Texture Colour
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 22 / 34
Results
Texture Color0
0.5
1
1.5
2
2.5
d’
S1
Texture Color0
0.5
1
1.5
2
2.5
d’
S2
Texture Color0
0.5
1
1.5
2
2.5
d’
S3
Texture Colour Combined
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 23 / 34
Results
Texture Color Combined0
0.5
1
1.5
2
2.5
d’
S1
Texture Color Combined0
0.5
1
1.5
2
2.5
d’
S2
Texture Color Combined0
0.5
1
1.5
2
2.5
d’
S3
Texture Colour Combined
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 24 / 34
Stimulus space:
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 25 / 34
Orientationcontrast
Colour contrast
Stimulus space: Cue-conflict condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 26 / 34
Orientationcontrast
Colour contrast
Stimulus space: Cue-conflict condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 27 / 34
Orientationcontrast
Colour contrast
Stimulus space: Cue-conflict condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 28 / 34
Orientationcontrast
Colour contrast
Decision space: Cue-conflict condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 29 / 34
Internalresponse:Orientationcontrast
Internal response: Colour contrast
Decision space: Cue-conflict condition
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 30 / 34
Internalresponse:Orientationcontrast
Internal response: Colour contrast
Results
Texture Color Combined0
0.5
1
1.5
2
2.5
d’
S1
Texture Color Combined0
0.5
1
1.5
2
2.5
d’
S2
Texture Color Combined0
0.5
1
1.5
2
2.5
d’
S3
Texture Colour Combined Conflict
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 31 / 34
Results
Texture Color Combined Conflict0
0.5
1
1.5
2
2.5
d’
S1
Texture Color Combined Conflict0
0.5
1
1.5
2
2.5
d’
S2
Texture Color Combined Conflict0
0.5
1
1.5
2
2.5
d’
S3
Texture Colour Combined Conflict
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 32 / 34
Summary
1. Same information, different performance
d′
combined > d′
single cue ≥ d′
conflict
Although the cue-combination and cue-conflict stimuli had the same
texture and colour differences.
⇒ Integration is possible only with spatially aligned cues.
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 33 / 34
Summary
1. Same information, different performance
d′
combined > d′
single cue ≥ d′
conflict
Although the cue-combination and cue-conflict stimuli had the same
texture and colour differences.
⇒ Integration is possible only with spatially aligned cues.
2. Better cue-combination performance than predicted
d′
combined >
√
d′2texture + d
′2colour
Performance with combined cues is better than the best
independent-processing prediction.
⇒ Texture and colour are not processed independently.
⇒ Cue-invariant segmentation
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 33 / 34
Thank you!
Supported by:
Swiss National Science Foundation PBELP1-125415
NIH EY16165
Saarela & Landy (NYU) Visual segmentation ECVP 2010-08-24 34 / 34