14
Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm from the vision community Performance Multiway graph cut problem (NP-hard) Applications to vision problems Approximation algorithms from the theory and vision communities.

Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Embed Size (px)

DESCRIPTION

Image segmentation Image segmentation separates an image into disjoint objects. Olga Veksler proposed using minimum graph cuts to segment an image into a hierarchy of nested regions. Any pair of regions will either be disjoint or one region will be entirely contained within the other.

Citation preview

Page 1: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Outline Standard 2-way minimum graph cut problem.

Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm from the vision community Performance

Multiway graph cut problem (NP-hard) Applications to vision problems Approximation algorithms from the theory and

vision communities.

Page 2: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Minimum Cut Problem Separate the nodes

into two set, one containing s and one containing t.

Minimize the sum of the weights of the edges crossing the partition.

Equivalent to finding the maximum flow.

s

t

Page 3: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Image segmentation Image segmentation separates an

image into disjoint objects. Olga Veksler proposed using minimum

graph cuts to segment an image into a hierarchy of nested regions.

Any pair of regions will either be disjoint or one region will be entirely contained within the other.

Page 4: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Each pixel is a node, edges between neighbors

Edges weighted by difference in pixel colors

Border pixels link to Sink Compute the min-cut around

every pixel! Recurse on subgraphs to get

nested segments Connect border to a sink

Nested Cuts for Images

…SourceSink

Page 5: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Nested Cuts Image Results

Page 6: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Augmenting Paths algorithms Repeatedly search the residual graph for a

source-to-sink path. Send as much flow as possible down the path (saturate the minimum residual capacity edge).

How do we choose the path to augment? Largest Capacity/Capacity Scaling (O(m2 log U)) Edmunds and Karp - Shortest Path (Breadth First

Search) (O(nm2)) Dinic’s Method - Layered networks (O(n2m))

Page 7: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Push Relabel Algorithms Initially push as much flow as possible to all

nodes connected to the source. Pick a node u with excess flow. Pick an edge (u,v) with excess capacity

where v has a label which is one less than u. Push as much flow as possible to v.

If not possible, increase the label of u. Rinse and repeat. Think of labels as the distance from the sink.

Page 8: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Example

4

2 1

3/51/6

L=2

L=4 L=1

2

2 3

5/51/6

L=4 L=1

L=2

4

2 1

3/51/6

L=2

L=5 L=3

4

2 1

3/51/6

L=5 L=3

L=4

Push:

Relabel:

Page 9: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Choosing the next vertex FIFO queue: O(n3) Highest label node - maintain an array

of sets where all nodes in a given set have the same label:

Largest excess: O(n2m) Stack: O(n2m)

O(n2 m )

Page 10: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Relabeling heuristics Global relabeling: periodically relabel every

node with the shortest path distance to the sink in the residual graph.

Gap relabeling: Periodically check for gaps in the labels. If a gap is found, relabel every node above the gap with n (# of vertices), because the sink is not reachable from these nodes.

Page 11: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Boykov-Kolmogorov algorithm

An augmenting paths algorithm from the vision literature, designed to be efficient on the graphs that typically arise in vision problems. O(n2m|C|)

Maintains a tree (rooted at the source) of nodes connected by edges with excess capacity.

The tree is grown until the sink is part of the tree.

Page 12: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Boykov-Kolmogorov, cont. We then augment down the path to the

sink . This may split the tree into a forest

(because edges are saturated): a tree containing the root and some orphans. Attempt to adopt the orphan trees.

Repeat until the growth phase is unable to incorporate the sink.

Page 13: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Performance Experimental comparisons:

Cheung showed that Dinic’s is the fastest augmenting paths algorithm.

Anderson and Setubal showed that Push Relabel is faster than Augmenting Paths, and that the FIFO queue and Highest Label methods are the fastest.

Page 14: Outline Standard 2-way minimum graph cut problem. Applications to problems in computer vision Classical algorithms from the theory literature A new algorithm

Performance More experimental comparisons:

Cherkassky and Goldberg showed that the quickest push relabel method is Highest Label with both gap and global relabeling.

Boykov and Kolmogorov showed that their algorithm is the fastest on the range of vision problems they tested it on:

Image restoration Stereo with occlusions Interactive object segmentation