![Page 1: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/1.jpg)
On Measuring* the Ecological Validity of Local Figure-Ground Cues
Charless Fowlkes, David Martin, Jitendra Malik
Computer Science DivisionUniversity of California at Berkeley
* "When you can measure what you are speaking about and express it in numbers, you know something about it; but when you
cannot measure it, when you cannot express it in numbers, your knowledge is of the meager and unsatisfactory kind." --Lord Kelvin
![Page 2: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/2.jpg)
2
Cues to Figure-Ground Assignment
• Size
• Surroundedness
• Convexity
• Lower-Region
• Symmetry
• Parallelism
• Meaningfulness
Photo by Wei-Chung Lee
![Page 3: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/3.jpg)
3
Ecological Statistics of Figure-Ground Cues
• Hypothesis: Perceptual organization reflects the statistics of the natural world in which the visual system evolved.
• In the context of grouping, this has been explored by:– Brunswik/Kamiya 1953 : proximity of similars
– Geisler et. al. 2001 : good continuation
– Martin/Fowlkes/Malik 2001 : proximity, similarity in color/texture
• In this work we measure, in a probabilistic sense, the power of size, convexity and lower-region in determining figure-ground assignment
![Page 4: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/4.jpg)
4
1. Human observers assign figure-ground labels to every boundary in a collection of natural images.
2. The cues of size, convexity, and lower-region are measured locally at each boundary point.
3. The extent to which these local cues are able to predict the ground-truth labeling is quantified.
Overview
![Page 5: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/5.jpg)
5
Berkeley Segmentation Dataset
1000 images each segmented by 10 different subjectshttp://www.cs.berkeley.edu/projects/vision/grouping/segbench
![Page 6: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/6.jpg)
6
Figure-Ground Labeling
- 200 segmented images of natural scenes- boundaries labeled by at least 2 different human subjects- subjects agree on 88% of contours labeled
![Page 7: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/7.jpg)
7Size(p) = log(AreaF / AreaG)
Size and Surroundedness [Rubin 1921]
GFp
![Page 8: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/8.jpg)
8
Convexity(p) = log(ConvF / ConvG)
Convexity [Metzger 1953, Kanizsa and Gerbino 1976]
ConvG = percentage of straight lines that lie completely within region G
pG F
![Page 9: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/9.jpg)
9
LowerRegion(p) = θG
Lower Region[Vecera, Vogel & Woodman 2002]
θ
p
center of mass
![Page 10: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/10.jpg)
10
size convexity lower region
-Sample 350,000 boundary points from 200 images -Intersect with circular window of chosen radius r-Compute size, convexity and lower-region cues and compare to ground truth labeling
![Page 11: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/11.jpg)
11
Figural regions tend to be smaller
mean is zero with p < 10-16
![Page 12: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/12.jpg)
12
Figural regions tend to be convex
mean is zero with p = 0.021 (less at other radii)
![Page 13: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/13.jpg)
13
Figural regions tend to lie below ground regions
mean is 90 with p < 10-16
![Page 14: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/14.jpg)
14
Power of cue depends on support of the analysis window.
![Page 15: On Measuring * the Ecological Validity of Local Figure-Ground Cues Charless Fowlkes, David Martin, Jitendra Malik Computer Science Division University](https://reader035.vdocument.in/reader035/viewer/2022062714/56649d615503460f94a42e03/html5/thumbnails/15.jpg)
15
Conclusion
• Figural regions are smaller, more convex and below ground regions in natural images