Multiple Random Walkers and Their Application to Image Cosegmentation
Chulwoo
1
Lee ,
Won-Dong
1Korea
Contents
1. Multiple Random Walkers (MRW)
2. MRW with Interaction: Repulsive Restart Rule
3. Application to Image Cosegmentation
1
Jang ,
Jae-Young
2
Sim ,
and Chang-Su
1
Kim
University, Seoul, Korea, 2UNIST, Ulsan, Korea
Image Cosegmentation
• Repulsive restart rule for MRW clustering
• Inter-image transfer matrix
− SIFT, LAB, and texton
• Decision by the MAP rule
• Foreground estimation by RWR
Image Examples
Multiple Random Walkers
• Multi-Pass Refinement
• Single Random Walker
Stationary distribution
• Single Random Walker with Restart (RWR)
• Multiple Random Walkers (MRW)
− Interactions by time-varying restart distribution
MRW with Repulsive Restart Rule
• Image to graph
− Super-pixel nodes
− Boundary connected and extended edges
• Feature distances for the edge weight
− Average RGB and LAB
− Boundary cue
− Bag-of-visual-word: RGB and LAB
(1)
(2)
(3)
(4)
1) Initials with the center priors
2) Inter-image concurrence computation
• Double random walkers
• Similarity of each node in image
to foreground objects in the other images
3) Intra-image MRW clustering
• Designing restart rule
for clustering
• Agent should have large restart at the dominant nodes – posterior
• Different characteristics
Ncut
• Agent should have smooth probability transition in the recursion – current
probability
SKM
No initial prior
RWR
MRW
Fixed
Interactive
Divide images along the strongest edges Dependent on initials Meaningful foregrounds
Hybrid restart rule
Repulsive
(Interactive)
Concurrence
(Fixed)
4) Repeat until the overall foreground distance stops decreasing