Fast, Multiscale Image Fast, Multiscale Image Segmentation: From Pixels to Segmentation: From Pixels to Semantics Semantics Ronen Basri The Weizmann Institute of Science Joint work with Achi Brandt, Meirav Galun, Eitan Sharon
Dec 19, 2015
Fast, Multiscale Image Fast, Multiscale Image Segmentation: From Pixels to Segmentation: From Pixels to
SemanticsSemantics
Ronen BasriThe Weizmann Institute of Science
Joint work with
Achi Brandt, Meirav Galun, Eitan Sharon
SSegmentation by egmentation by WWeighted eighted AAggregationggregation
A multiscale algorithm:• Optimizes a global measure• Returns a full hierarchy of segments• Linear complexity• Combines multiscale measurements:
– Texture– Boundary integrity
The Pixel GraphThe Pixel GraphCouplings (weights)
reflect intensity
similarity
Low contrast –strong couplingHigh contrast –weak coupling
i jI I
ijW e
Normalized-cut MeasureNormalized-cut Measure
( )( )
( )
Cut SE S
Int S
Minimize:
2( ) ( )ij i ji j
Cut S w u u
( ) ij i ji j
Int S w u u
Si
Siui 0
1
Saliency MeasureSaliency Measure
2( ) ( )
1( )
2
( ) 2
Tij i j
i j
Tij i j
i j
T
T
Cut S w u u u Lu
Int S w u u u Wu
u LuE S
u Wu
Lu WuMinimize:
Multiscale Computation Multiscale Computation of Ncutsof Ncuts
• Our objective is to rapidly find the segments (0-1 partitions) that optimize
• For single-node cuts we simply evaluate • For multiple-node cuts we perform “soft
contraction” using coarsening procedures from algebraic multigrid solvers of PDEs.
Coarsening the GraphCoarsening the Graph
• Suppose we can define a sparse mappingsuch that for all minimal states
: , ( / 2)N nP N n R R
11
22
.
..
Nn
Uu
P
uU
uU
Coarse EnergyCoarse Energy
• Then
• PTWP, PTLP define a new (smaller)
graph
( ) 2 2T T T
T T T
u Lu U P LPUE S
u Wu U P WPU
Recursive CoarseningRecursive Coarsening
11
22
.
..
Nn
Uu
P
uU
uU
For a salient segment :
( )n NP ,sparse interpolation matrix
iu julUkU
Weighted AggregationWeighted Aggregation
ijwi
jjlp
aggregate k aggregate l
[[ 1] ]s T sWW P P
klWikp
HierarchicHierarchical Graphal Graph
Pyramid of graphs Soft relations between levels Segments emerge as salient nodes at some level of the pyramid
Physical MotivationPhysical Motivation
• Our algorithm is motivated by algebraic multigrid solutions to heat or electric networks
• u - temperature/potential• a(x, y) – conductivity• At steady state largest temperature
differences are along the cuts• AMG coarsening is independent of f
( , ) ( , ) ( , )u u
a x y a x y f x yx x y y
A Chicken and Egg A Chicken and Egg Problem Problem
Problem:Coarse measurements mix neighboring statistics
Solution: Support of measurements is determined as the segmentation process proceeds
Hey, I was here first
Texture AggregationTexture Aggregation
• Aggregates assumed to capture texture elements
• Compare neighboring aggregates according to the following statistics:– Multiscale brightness measures– Multiscale shape measures– Filter responses
• Use statistics to modify couplings
Recursive Computation of Recursive Computation of MeasuresMeasures
• Given some measure of aggregates at a certain level (e.g., orientation)
• At every coarser level we take a weighted sum of this measure from previous level
• The result can be used to compute the average, variance or histogram of the measure
• Complexity is linear
Adaptive vs. Rigid Adaptive vs. Rigid MeasurementsMeasurements
Averaging
Our algorithm - SWA
Original
Geometric
Adaptive vs. Rigid Adaptive vs. Rigid MeasurementsMeasurements
Interpolation Geometric
Original
Our algorithm - SWA
Key DifferencesKey Differences
• Optimize a global measure(like Malik’s Ncuts)
• Hierarchy with soft relations(unlike agglomerative/graph contraction)
• Combine texture measurements while avoiding the “chicken and egg problem”
ComplexityComplexity
• Every level contains about half the nodes of the previous level:
Total #nodes 2 X #pixels• All connections are local, cleaning small
weights• Top-down sharpening: constant number
of levels• Linear complexity• Implementation: 5 seconds for 400x400
MS Lesion DetectionMS Lesion Detection
TaggedTagged Our resultsOur results
Data: FilippiData: Filippi