P. Rodríguez, R. Dosil, X. M. Pardo, V. LeboránP. Rodríguez, R. Dosil, X. M. Pardo, V. Leborán
Grupo de Visión ArtificialGrupo de Visión ArtificialDepartamento de Electrónica e ComputaciónDepartamento de Electrónica e Computación
Universidade de Santiago de CompostelaUniversidade de Santiago de Compostela
Automatic Generation ofAutomatic Generation ofInitial Surfaces for Implicit SnakesInitial Surfaces for Implicit Snakes
IntroductionIntroduction Global Shape ModelGlobal Shape Model
CSG ModelCSG Model Superquadric primitivesSuperquadric primitives
MethodologyMethodology Prior Model ConstructionPrior Model Construction Image Feature Extraction Image Feature Extraction MatchingMatching
Results and ConclusionsResults and Conclusions
OutlineOutline
IntroductionIntroduction 33D surface reconstruction:D surface reconstruction:
Segmentation with deformable modelsSegmentation with deformable models
Good local approximationGood local approximation
Need of good initial estimationNeed of good initial estimation
IntroductionIntroduction Previous SolutionsPrevious Solutions
Manual initialization: Manual initialization: is not practical in 3Dis not practical in 3D
Landmark registration: Landmark registration: landmarks are not always identifiablelandmarks are not always identifiable
Part decomposition techniquesPart decomposition techniques need of joint detection or part recoveryneed of joint detection or part recovery lack of robustness when data is incomplete or noisylack of robustness when data is incomplete or noisy
IntroductionIntroduction ObjectivesObjectives
Automatic initialization of 3D medical imagesAutomatic initialization of 3D medical images(CT, MRI, …)(CT, MRI, …)
No use of landmarksNo use of landmarks
Application to multi-part objectsApplication to multi-part objects
Robustness to noise and presence of other Robustness to noise and presence of other objectsobjects
IntroductionIntroduction Proposal:Proposal:
matching with multi-part prior modelsmatching with multi-part prior models
Initialization by matching with prior modelsInitialization by matching with prior models RobustnessRobustness No need of part or joint detectionNo need of part or joint detection
Use of composite global shape modelsUse of composite global shape models Multi-part models: CSGMulti-part models: CSG Primitives: SuperquadricsPrimitives: Superquadrics
Image features are image surface pointsImage features are image surface points No use of landmarksNo use of landmarks
Average Surface
Prior Model
I. Modeling
I. Prior model construction from sample images
Volume Data
Surface Patches
II. Preprocessing
II. Object surface points extraction
III. Matching
Initial Model
III. Matching between surface model and object surface points
IntroductionIntroduction
Global Shape ModelGlobal Shape Model Constructive Solid Geometry (CSG)Constructive Solid Geometry (CSG)
Binary treeBinary tree Leaf nodes: solid primitivesLeaf nodes: solid primitives Internal nodes: Boolean operationsInternal nodes: Boolean operations Arcs: rigid transformationsArcs: rigid transformations
Primitives: Primitives: Superquadrics with global deformationsSuperquadrics with global deformations
Global Shape ModelGlobal Shape Model Superquadrics with global deformationsSuperquadrics with global deformations
Few parameters bring structural informationFew parameters bring structural information
Global Deformations: asymmetryGlobal Deformations: asymmetry
Implicit equationImplicit equation
1, qrf
MethodologyMethodologyI. Prior model construction from
sample images
Manual part decompositionManual part decomposition
Individual modeling of object partsIndividual modeling of object parts Shape parametersShape parameters Relative spatial distribution Relative spatial distribution
parametersparameters
Mqq ,...,1m
Average Surface
Prior Model
I. Modeling
siii
tisii
T qrqr
qqq
,
MethodologyMethodologyI. Prior model construction from
sample images
Optimization with Genetic Optimization with Genetic AlgorithmsAlgorithms
Minimization of error function:Minimization of error function:
where where
andand
N
iiDE
1
22 ,, qrq x
Average Surface
Prior Model
I. Modeling
qr
rqr,
11,
1εfD
Nrr ,...,1x
MethodologyMethodologyII. Image feature extraction
1.1. Smoothing by anisotropic Smoothing by anisotropic diffusiondiffusion
2.2. Non gradient maxima Non gradient maxima suppressionsuppression
3.3. Hysteresis thresholdingHysteresis thresholding
Volume Data
Surface Patches
II. Preprocessing
MethodologyMethodologyIII. Matching between model and object features
Find global rigid transformation Find global rigid transformation TT such that the such that the transformed model fits the object transformed model fits the object surfacesurface
GA to minimize error functionGA to minimize error function
N
iji
jD
NE
1
22 ',min1
, qrxm'
Surface Patches
III. Matching
Initial Model
Prior Model
',...,'1 Mqqm'
MethodologyMethodologyIII. Matching between model and object features
Radial distance to a Radial distance to a deformeddeformed implicit surface implicit surface
is difficult to calculateis difficult to calculate The following approximation is usedThe following approximation is used
sjijTTD qr ,11
', jiD qr
ResultsResults
ResultsResults
ConclusionsConclusions ContributionsContributions
Automatization of initializationAutomatization of initialization Easy handling of multipart shapes using a compound Easy handling of multipart shapes using a compound
modelmodel No part or joint detection No part or joint detection Easy optimization of the modelEasy optimization of the model
Future workFuture work Introduction of fine tuning of individual part Introduction of fine tuning of individual part
parametersparameters Incorporation of other Boolean operations to the CSG Incorporation of other Boolean operations to the CSG
model to handle concavitiesmodel to handle concavities