A Co-Evaluation Framework for Improving Segmentation Evaluation

A CoA Co--Evaluation Framework for Evaluation Framework for

Improving Segmentation EvaluationImproving Segmentation Evaluation

Hui Zhang, Jason Fritts, and Sally Goldman

Washington University in St. Louis

Department of Computer Science and Engineering

� Image Segmentation & Evaluation

� Co-Evaluation Framework

� Experiments

� Human vs. Machine Segmentation

� Machine vs. Machine Segmentation

� Conclusions & Future Work

Outline

� Partitioning an image into separate regions

Image Segmentation

� A critical early step towards content analysis and image understanding

A B C

� Three categories of segmentation methods

� Pixel-based

� Histogram thresholding, etc

Image Segmentation

� Region-based� Split-and-merge, region growing, etc

� Edge-based� Edge flow, color snakes, etc

� Problems

� Unable to compare different segmentation

methods, or even different parameterizations of a

single method

� Unable to determine whether one segmentation

method is appropriate for different types of images

� Segmentation evaluation is of critical

importance

Image Segmentation

Segmentation Evaluation

Original image

HS

IHS

JSEG

Image-

Magick

Edison

eCognition

Original image

Improved Hierarchical Segmentation (IHS)

10 15 20

Different parameterizations

Of a single method:

Segmentation Evaluation

Image

Segmentation

evaluationObjective

evaluation

Subjective

evaluation

System-level

evaluation

Direct

evaluation

Analytical

methods

Empirical

methods

Relative

methods

Stand-alone

methods

[ Taxonomy ]

Segmentation Evaluation

� Relative Evaluation (a.k.a. Empirical Discrepancy Measures)

� Most widely used objective evaluation methods

� Comparing results to manually-segmented reference

image

Original image reference image Result from EDISON

Measure the discrepancy

Objective Segmentation Evaluation

� Stand-alone Evaluation (a.k.a. empirical goodness measures)� No reference image is needed

� Examine how well it matches the desired characteristics of segmented images, based on human intuition

� A goodness score is given for each segmentation result

Segmentation

Result 1

Result 2 from

EDISON

Stand-alone Objective Evaluation

Stand-alone Evaluation

Goodness Score: 6.1345 Goodness Score: 8.2733

� Intra-region uniformity measures

� Color variance

� Squared color error

� Texture

� Inter-region disparity measures

� Average color difference between regions

� Average color difference along boundary

� Semantic measures

� Shape

Stand-alone Objective Evaluation

� Liu and Yang’s F function:

� Borsotti et al.’s Q function:

� Correia and Pereira’s V Functions:

Stand-Alone Objective Evaluation

Functions

contrastcompactelongcircRvs 55.225._225.)( ++=

contrastcompactelongcircRvm 40.30._30.)( ++=

( )∑

=

+

+×=

N

j j

j

j

j

I S

SN

S

eN

SIQ

1

22

log11000

1)(

∑=

=N

i i

i

S

eNIF

1

2

1000

1)(

� Utilize minimum description length (MDL) principle

� regards segmentation as a modeling process

� MDL provides effective balance between:

� uniformity within a region

� complexity of segmentation

� Our evaluation function, E, based on MDL:

Evaluation function E

)()()|()( IHIHSRLSLE rl +=+=

Evaluation function E

Region entropy

Layout entropy

E

Original image

Segmentation 1ayout

� Image Segmentation & Evaluation

� Co-Evaluation Framework

� Experiments

� Human vs. Machine Segmentation

� Machine vs. Machine Segmentation

� Conclusions & Future Work

Outline

Co-Evaluation framework

Co-Evaluation framework

Evaluator 1

Evaluator 2

Evaluator N

Combiner

Original image

Segmented images

12

Simple combiner strategy:

111221result

combinerEvaluator 5Evaluator 4Evaluator 3Evaluator 2Evaluator 1

Example:

Co-Evaluation framework

Evaluator 1

Evaluator 2

Evaluator N

Combiner

Original image

Segmented images

12

Human Evaluation Results: For Training

Expert combiner strategy:

Example:

211121evaluation

211221training

combinerHumanEvaluator5Evaluator4Evaluator3Evaluator2Evaluator1

� Image Segmentation & Evaluation

� Co-Evaluation Framework

� Experiments

� Human vs. Machine Segmentation

� Machine vs. Machine Segmentation

� Conclusions & Future Work

Outline

Human vs. Machine

Segmentation Evaluation

� 5 evaluators: E, F, Q, Vs, Vm

� Image pair (human segmentation, machine segmentation)

� 108 training sets from Berkeley dataset and Military Graphics Collection (aircraft)

Human vs. Machine

Segmentation Evaluation Results

� 108 training pairs

� Results for 92 evaluation pairs:

Single evaluation method:

Co-Evaluation:

� Image Segmentation & Evaluation

� Co-Evaluation Framework

� Experiments

� Human vs. Machine Segmentation

� Machine vs. Machine Segmentation

� Conclusions & Future Work

Outline

Machine vs. Machine

Segmentation Evaluation Results

� 6 subjective evaluators determined 3 best and 3 worst segmentations across image regressions of 2 to 20 segments

� Most common best and worst segmentations from each image used toform Best set and Worst set

� 125 pairs of (Best, Worst) used for training

� 124 pairs used for evaluation

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

E F Q Vs Vm SC WM NB SVM

Ac

cu

rac

y

Conclusions

� By themselves,

� E is best at determining human vs. machine segmentations

� Q is best at comparing machine vs. machine segmentations

� Using the Co-Evaluation framework,

� Naive Bayesian combiner does best

� Weighted Majority and Support Vector Machines are also

good, much better than a simple combiner

� Some gain from Co-Evaluation framework, but

expect we can do better…

Future Work… Meta-Learning

Evaluator 1

Evaluator 2

Evaluator N

Combiner

Original image

Segmented images

12

Human Evaluation Results: For Training

Images

features

Meta-learner strategy:

2textured11211evaluation

2textured11212Training

combinerhumanMeta-featureEval.5Eval.4Eval.3Eval.2Eval.1

Example:

Thank You!

� More information:

� http://www.cse.wustl.edu/~jefritts/