boundary extraction in natural images using ultrametric contour maps

36
Boundary Extraction in Natural Images Using Ultrametric Contour Maps Pablo Arbeláez Université Paris Dauphine Presented by Derek Hoiem

Upload: isaiah

Post on 07-Feb-2016

39 views

Category:

Documents


0 download

DESCRIPTION

Boundary Extraction in Natural Images Using Ultrametric Contour Maps. Pablo Arbel á ez Universit é Paris Dauphine Presented by Derek Hoiem. What is segmentation?. What is segmentation?. Segmentation is a result. Face. Woman. What is segmentation?. Segmentation is a result - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Pablo Arbeláez Université Paris Dauphine

Presented byDerek Hoiem

Page 2: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

What is segmentation?

Page 3: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

What is segmentation?

• Segmentation is a result

Page 4: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

What is segmentation?

• Segmentation is a result

• Segmentation is a process

Woman

Face

Page 5: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

What is segmentation?

• Segmentation is a result

• Segmentation is a process

• Segmentation is a guide

Page 6: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Segmentation as a Guide

• Multiple Segmentations

Page 7: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Segmentation as a Guide

• Multiple Segmentations

• Hierarchy of Segmentations

Page 8: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Key Concepts/Contributions

• Hierarchical segmentation by iterative merging

• Ultrametric dissimilarities

• Thorough evaluation on BSDS

Page 9: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Hierarchical Segmentation

λ

3 Region Image Dendrogram

Contour Image

Page 10: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Ultrametric Contour Map

• Ultrametric– Definition: D(x,y) <= max{ D(x,z), D(z,y) }

The union R12 of two regions R1 and R2 must have >= distance to adjacent region R3 than either R1 or R2

λ

Page 11: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Ultrametric Contour Map

Page 12: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Region Dissimilarity

1. Dc(R1, R2): mean boundary contrast– contrast(x) = max L*a*b* diff within radius of x

2. Dg(R1, R2): mean boundary gradient– gradient(x) = Pb(x)

3. Da(R1): Area + α3 Scatter (in color space)

D(R1, R2) = [Dc(R1, R2) + α1 Dg(R1, R2)] · min{ Da(R1) , Da(R2) }α2

Learned Parameters: xi = 4.5 α1 = 5 α2 = 0.2 α3 = 0

Page 13: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Examples

Contrast

Contrast + Gradient

Contrast + Gradient + Region

Page 14: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Algorithm Summary

• Create Initial Contours:– Extrema in gray channel form regions– Assign pixels to regions based on above

ultrametric

• Iteratively merge regions– Keep adjacency/distance matrix

Page 15: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Comparison

• Martin et al. (Pb)• Canny edge detector• Hierarchical watersheds (using MFM for gradient)

[Najman and Schmitt 1996]• Variational (global energy minimization)

Page 16: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Pb

No Boundary

Boundary

[Martin Fowlkes Malik 2004]

Oriented Edges

Brightness Gradient

Color Gradient

Texture Gradient

Page 17: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Pb

Page 18: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Variational Method

[Koepfler Lopez Morel 1994]

Originally Wavelet-based Textons

Page 19: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Comparison• MFM: Martin et al. (Pb)• Canny: Canny edge detector• WS: Hierarchical watersheds (using MFM for gradient) [Najman and Schmitt 1996]• MS: Variational (global energy minimization)

Edge-Based Region-Based

Page 20: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Comparison

Page 21: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Results

Page 22: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Results

Page 23: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Best Results

http://www.ceremade.dauphine.fr/~arbelaez/results-UCM/main.html

Page 24: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Best Results

http://www.ceremade.dauphine.fr/~arbelaez/results-UCM/main.html

Page 25: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Best Results

http://www.ceremade.dauphine.fr/~arbelaez/results-UCM/main.html

Page 26: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Best Results

http://www.ceremade.dauphine.fr/~arbelaez/results-UCM/main.html

Page 27: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Median Results

Page 28: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Median Results

Page 29: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Median Results

Page 30: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Median Results

Page 31: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Worst Results

Page 32: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Worst Results

Page 33: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Worst Results

Page 34: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Worst Results

Page 35: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Hierarchies vs. Multiple Segmentations

Page 36: Boundary Extraction in Natural Images Using Ultrametric Contour Maps

Revising Segmentation