for Automatic Object Segmentation SasiKanth Bendapudi Yogeshwar Nagaraj What is a Good Segmentation http wwweecsberkeleyedu Research Projects CSvisiongrouping resourceshtml ID: 919335
Download Presentation The PPT/PDF document "Constrained Parametric Min-Cuts" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Constrained Parametric Min-Cuts for Automatic Object Segmentation
SasiKanth
Bendapudi
Yogeshwar
Nagaraj
Slide2What is a ‘Good Segmentation’?
Slide3http://
www.eecs.berkeley.edu
/Research/
Projects
/CS/vision/grouping/
resources.html
Slide4“Geometric context from a single image”,
Hoiem
et al.
, ICCV 2005
Slide5“Using
Multiple Segmentations to Discover
Objects and
their Extent in Image
Collections”,
Russel
et al.
, CVPR 2006
Slide6“Improving
Spatial Support for Objects via Multiple
Segmentations”,
Malisiewicz
&
Efros
, BMVC 2007
Slide7“Towards
Unsupervised Whole-Object
Segmentation: Combining
Automated Matting with Boundary
Detection”, Stein & Hebert, CVPR 2008
Slide8The PaperFigure-Ground segmentationSolve CPMC by minimizing the objective function using various seeds and parameters
Reject redundancies and obvious negatives based on segment energies and similarities
Learn the characteristics of a ‘Figure’ segment to qualitatively assess the remaining segments
Slide9Objective Function
Slide10Objective Function
Slide11Objective Function
Slide12Synthetic Example
Slide13Synthetic Example
Slide14InitializationForeground
Regular 5x5 grid geometry
Centroids of large N-Cuts regions
Centroids of
superpixels
closest to grid positions
Background
Full image boundaryHorizontal boundaries
Vertical boundariesAll boundaries excluding the bottom one
Performance broadly invariant to different initializations
Slide15Fast Rejection
Large set of initial segmentations (
~5500
)
High Energy
Low Energy
~
2000
segments with the lowest energy
Cluster segments based on spatial
overlap
L
owest
energy member of each cluster (
~154
)
Slide16Segment RankingModel data using a host of features
Graph partition properties
Region properties
Gestalt properties
Train a
regressor
with the largest overlap ground-truth segment using
Random ForestsDiversify similar rankings using Maximal Marginal Relevance (MMR)
Slide17Graph Partition PropertiesCut – Sum of affinities along segment boundary
Ratio Cut – Sum along boundary divided by the number
Normalized Cut – Sum of cut and affinity in foreground and background
Unbalanced N-cut – N-cut divided by foreground affinity
Thresholded
boundary fraction of a cut
Slide18Region Properties
Area
Perimeter
Relative Centroid
Bounding Box properties
Fitting Ellipse properties
Eccentricity
Orientation
Convex
Area
Euler Number
Diameter of Circle with the same area of the segment
Percentage of bounding box covered
Absolute distance to the center of the
image
Slide19Gestalt PropertiesInter-region
texton
similarity
Intra-region
texton
similarity
Inter-region brightness similarity
Intra-region brightness similarityInter-region contour energy Intra-region contour energy
Curvilinear continuityConvexity – Ratio of foreground area to convex hull area
Slide20Feature Importance
Slide21Feature Importance
Slide22Feature Importance
Slide23What has been modeled?
Slide24DatabasesWeizmann databaseF-measure criterion
MSR-Cambridge database & Pascal VOC2009
Segmentation covering
Slide25Performance
Slide26Test of the algorithmBerkeley segmentation datasetComplete pool of images collected
Ranked using the ranking methodology
Top ranks evaluated to test the ranking procedure
How well does the algorithm perform?
Slide27Berkeley Database
Rank 269!
Slide28Berkeley Database
Rank 142!
Slide29Berkeley Database
Rank 98!
Slide30Berkeley DatabaseCompute the Segment Covering score for the top 40 segments of each image in the database
Database
Segment Covering Score (Top 40)
BSDS
0.52
MSR Cambridge
0.77
Pascal VOC
0.63
Database
Segment Covering Score
(All segments)
BSDS
0.61
MSR Cambridge
0.85
Pascal VOC
0.78
Slide31ConclusionDoes Constrained Parametric Min-Cuts
work well?
Yes
Does
Fast Rejection
work well?
Yes
Does
Segment Ranking work well?
I don’t think so
Slide32Interesting follow up
Image
Segmentation by Figure-Ground Composition into Maximal
Clique
,
Ion,
Carreira
,
Sminchisescu, ICCV 2011
Slide33Interesting follow up
Image
Segmentation by Figure-Ground Composition into Maximal
Clique
,
Ion,
Carreira
, Sminchisescu, ICCV 2011
Obtain pool of FG segmentations from CPMCDefine
tiling
and a probabilistic model for the same
Represent the probabilistic models using mid-level features
Compute and rank various
tilings
by implementing discrete searches from each of the nodes
Slide34Interesting follow up
Image
Segmentation by Figure-Ground Composition into Maximal
Clique
,
Ion,
Carreira
,
Sminchisescu, ICCV 2011
Slide35Questions?