SlideShare a Scribd company logo
Self-dependent 3D Face
Rotational Alignment Using
the Nose Region
Mehryar Emambakhsh
Adrian Evans
Department of Electronic and Electrical Engineering
University of Bath
Outline
 Introduction to 3D alignment
 SD 3D Face Rot Al
 Advantages of the nose region
 Algorithm steps
 Preprocessing and segmentation
 Nose region IFill
 Energy function minimisation
 Results
 Conclusions and future work
3
Introduction to 3D alignment
 3D face recognition algorithms
Holistic
Regional
Landmark-based
 3D object alignment errors
Scale
Translation
Rotation
4
Introduction to 3D alignment
 Alignment techniques
 A rotational invariant representation
 Extended Gaussian images
 Spherical harmonics
 Spin images
 Snapshot algorithm
 3D Fourier descriptors
 Registration-based algorithms
 Iterative closest point:
 Minimum square error
 Surface interpenetration measure
 Self-dependent algorithms
Advantages
• Generalisation
Disadvantages
• Computational
complexity
• A pre-aligned
reference image
required
• Local minima
5
Introduction to 3D alignment
 Self-dependent algorithms
Prior knowledge of the object of interest
 Geometry
 Curvature
Much faster than the other two methods
More robust for local minima
No reference image required
6
SD 3D Face Rot Al: Advantages of
the nose region
 Properties of the nose region
Hard to conceal without attracting suspicion
Easily detected
 Convexity
 Nose tip, usually the closest
point to the camera
Small variations over different expressions
7
SD 3D Face Rot Al: Preprocessing
and segmentation
 Outliers:
impulsive noise
 Fuzzy C-
means and
median filtering
 Missing data 
Morphological
filling and
interpolation
8
SD 3D Face Rot Al: Nose region
IFill
 Filling algorithm:
Applied to inverted 3D nose region
9
SD 3D Face Rot Al: Nose region
IFill
Morphological filling of the 2½D range image
10
SD 3D Face Rot Al: Nose region
IFill
Largest connected component of the filled
image found:
 Label with the highest frequency in the histogram.
11
SD 3D Face Rot Al: Energy
function minimisation
 Two energy (cost) functions
are introduced on the found
binary image domain:
 EExyxy to maximise the areato maximise the area
of the filled region:of the filled region:
 The 3D image is rotated aroundThe 3D image is rotated around
the X and Y axesthe X and Y axes
 The rotation procedure stops asThe rotation procedure stops as
the biggest area was foundthe biggest area was found
 Simulated annealing followed bySimulated annealing followed by
Levenberg-Marquart is used forLevenberg-Marquart is used for
function optimisationfunction optimisation
12
SD 3D Face Rot Al: Energy
function minimisation
 EEZZ to maximise theto maximise the
symmetry of the filledsymmetry of the filled
region:region:
 The 3D image is rotated aroundThe 3D image is rotated around
the Z axisthe Z axis
 The binary image is divided intoThe binary image is divided into
the left and right partsthe left and right parts
 The exclusive OR of isThe exclusive OR of is
calculatedcalculated
 The rotation procedure stopsThe rotation procedure stops
when highest symmetrywhen highest symmetry
obtainedobtained
13
SD 3D Face Rot Al: Energy
function minimisation
14
Results
 The Face Recognition Grand Challenge
(FRGC) is used for evaluation.
Including 557 subjects and 4947 range
images.
15
Results
 Exy minimisation
The binary map in the 1st
, 65th
, and final
iterations.
16
Results
 Ezminimisation
17
Results
 Consistency evaluationConsistency evaluation
 All of the images are first aligned using ourAll of the images are first aligned using our
algorithmalgorithm
 For each subject, each image in turn is usedFor each subject, each image in turn is used
as the reference and all other images of thatas the reference and all other images of that
subject aligned to it using ICP algorithmsubject aligned to it using ICP algorithm
 The final rotation matrix is compared with theThe final rotation matrix is compared with the
identity matrixidentity matrix
18
Results
 Average rotation matrix for all 4947 imagesAverage rotation matrix for all 4947 images
 LL22 norm of rot. matrixnorm of rot. matrix
19
Results
 Elapsed time
 Our
algorithm:
2.1671 ±
0.5773s
 Brute force
ICP
algorithm:
6.9064 ±
3.1815s
20
Results
 Performance over different expressions
21
Conclusions and future work
 A novel 3D face rotational alignment algorithm
has been proposed
 The algorithm capitalises on the nose region’s
consistency, good localisation and convex
features
 Our approach can be used for both fine and
coarse alignment purposes
 Using simulated annealing as a global
optimisation algorithm helps to avoid local
minima & the Levenberg-Marquart algorithm
performs the final tuning for B(x,y)
22
Conclusions and future work
 Future work
 The use of the final binary maps to improve for the
nose segmentation
 The use of the alignment algorithm for 3D face
recognition will be evaluated
 The final binary maps have the potential for better
landmark localisation for the nose region
 A more accurate computational complexity analysis is
required
Thank you!

More Related Content

PDF
Advanced Image Reconstruction Algorithms in MRIfor ISMRMversion finalll
PPTX
Normal Mapping / Computer Graphics - IK
PDF
Design and optimization of compact freeform lens array for laser beam splitti...
PDF
Leader Follower Formation Control of Ground Vehicles Using Dynamic Pixel Coun...
PPTX
Ray casting algorithm by mhm
PDF
Temporary Coherence 3D Animation
PDF
Evaluation of illumination uniformity metrics in design and optimization of ...
PDF
Optical modeling and design of freeform surfaces using anisotropic Radial Bas...
Advanced Image Reconstruction Algorithms in MRIfor ISMRMversion finalll
Normal Mapping / Computer Graphics - IK
Design and optimization of compact freeform lens array for laser beam splitti...
Leader Follower Formation Control of Ground Vehicles Using Dynamic Pixel Coun...
Ray casting algorithm by mhm
Temporary Coherence 3D Animation
Evaluation of illumination uniformity metrics in design and optimization of ...
Optical modeling and design of freeform surfaces using anisotropic Radial Bas...

What's hot (20)

PDF
Modeling and optimization of high index contrast gratings with aperiodic topo...
PPTX
Visual Saliency: Learning to Detect Salient Objects
PPSX
Three View Self Calibration and 3D Reconstruction
PPT
Vision Based Traffic Surveillance System
PDF
Photogrammetry - Space Resection by Collinearity Equations
PDF
ModelingOfUnsegmentedCloudPointData-RP-SanjayShukla
PPTX
Urban 3D Semantic Modelling Using Stereo Vision, ICRA 2013
PPT
mathpsy2012 poster_Shweta_3(1)
PPTX
Graphics
PDF
Atti Rinem20146 Tannino
PDF
Inverse kinematics
PPS
5 27pres
PPT
Parameter-free Modelling of 2D Shapes with Ellipses
PDF
Enhanced Tracking Aerial Image by Applying Fusion & Image Registration Technique
PDF
PR-240: Modulating Image Restoration with Continual Levels via Adaptive Featu...
PDF
Time Multiplexed VLSI Architecture for Real-Time Barrel Distortion Correction...
PPT
Automatic Dense Semantic Mapping From Visual Street-level Imagery
PDF
Single Image Fog Removal Based on Fusion Strategy
PPT
Miniproject final group 14
PPTX
raycasting. ppt
Modeling and optimization of high index contrast gratings with aperiodic topo...
Visual Saliency: Learning to Detect Salient Objects
Three View Self Calibration and 3D Reconstruction
Vision Based Traffic Surveillance System
Photogrammetry - Space Resection by Collinearity Equations
ModelingOfUnsegmentedCloudPointData-RP-SanjayShukla
Urban 3D Semantic Modelling Using Stereo Vision, ICRA 2013
mathpsy2012 poster_Shweta_3(1)
Graphics
Atti Rinem20146 Tannino
Inverse kinematics
5 27pres
Parameter-free Modelling of 2D Shapes with Ellipses
Enhanced Tracking Aerial Image by Applying Fusion & Image Registration Technique
PR-240: Modulating Image Restoration with Continual Levels via Adaptive Featu...
Time Multiplexed VLSI Architecture for Real-Time Barrel Distortion Correction...
Automatic Dense Semantic Mapping From Visual Street-level Imagery
Single Image Fog Removal Based on Fusion Strategy
Miniproject final group 14
raycasting. ppt
Ad

Similar to Self-dependent 3D face rotational alignment using the nose region (20)

PDF
Video Stitching using Improved RANSAC and SIFT
PDF
International Journal of Engineering Research and Development (IJERD)
PPT
Build Your Own 3D Scanner: Surface Reconstruction
PDF
Detecting Boundaries for Image Segmentation and Object Recognition
PDF
IJCER (www.ijceronline.com) International Journal of computational Engineerin...
PDF
Efficient 3D stereo vision stabilization for multi-camera viewpoints
PDF
Vision based non-invasive tool for facial swelling assessment
PPTX
Technical seminar
PPT
IGARSS11-Zhang.ppt
PDF
Secure System based on Dynamic Features of IRIS Recognition
PDF
Lecture 01 frank dellaert - 3 d reconstruction and mapping: a factor graph ...
PDF
Performance Evaluation of Illumination Invariant Face Recognition Algorthims
PDF
Two marks with answers ME6501 CAD
PDF
Machine learning for high-speed corner detection
PDF
Matching algorithm performance analysis for autocalibration method of stereo ...
PDF
MONOGENIC SCALE SPACE BASED REGION COVARIANCE MATRIX DESCRIPTOR FOR FACE RECO...
PDF
ME6501 Cad 2 mark question and answer
PDF
AN IMPROVED IRIS RECOGNITION SYSTEM BASED ON 2-D DCT AND HAMMING DISTANCE TEC...
PDF
Nose Tip Detection Using Shape index and Energy Effective for 3d Face Recogni...
Video Stitching using Improved RANSAC and SIFT
International Journal of Engineering Research and Development (IJERD)
Build Your Own 3D Scanner: Surface Reconstruction
Detecting Boundaries for Image Segmentation and Object Recognition
IJCER (www.ijceronline.com) International Journal of computational Engineerin...
Efficient 3D stereo vision stabilization for multi-camera viewpoints
Vision based non-invasive tool for facial swelling assessment
Technical seminar
IGARSS11-Zhang.ppt
Secure System based on Dynamic Features of IRIS Recognition
Lecture 01 frank dellaert - 3 d reconstruction and mapping: a factor graph ...
Performance Evaluation of Illumination Invariant Face Recognition Algorthims
Two marks with answers ME6501 CAD
Machine learning for high-speed corner detection
Matching algorithm performance analysis for autocalibration method of stereo ...
MONOGENIC SCALE SPACE BASED REGION COVARIANCE MATRIX DESCRIPTOR FOR FACE RECO...
ME6501 Cad 2 mark question and answer
AN IMPROVED IRIS RECOGNITION SYSTEM BASED ON 2-D DCT AND HAMMING DISTANCE TEC...
Nose Tip Detection Using Shape index and Energy Effective for 3d Face Recogni...
Ad

More from Mehryar (Mike) E., Ph.D. (7)

ODP
Deep Recurrent Neural Network for Multi-target Filtering
PDF
POL-LWIR Vehicle Detection: Convolutional Neural Networks Meet Polarised Infr...
PPT
Locating texture boundaries using a fast unsupervised approach based on clust...
PPT
Automatic MRI brain segmentation using local features, Self-Organizing Maps, ...
PPT
A Hybrid top-down/bottom-up approach for image segmentation incorporating col...
PPTX
An Evaluation of Denoising Algorithms for 3D Face Recognition
PPT
Using nasal curves matching for expression robust 3D nose recognition
Deep Recurrent Neural Network for Multi-target Filtering
POL-LWIR Vehicle Detection: Convolutional Neural Networks Meet Polarised Infr...
Locating texture boundaries using a fast unsupervised approach based on clust...
Automatic MRI brain segmentation using local features, Self-Organizing Maps, ...
A Hybrid top-down/bottom-up approach for image segmentation incorporating col...
An Evaluation of Denoising Algorithms for 3D Face Recognition
Using nasal curves matching for expression robust 3D nose recognition

Recently uploaded (20)

PPTX
Programs and apps: productivity, graphics, security and other tools
PDF
Blue Purple Modern Animated Computer Science Presentation.pdf.pdf
PDF
MIND Revenue Release Quarter 2 2025 Press Release
PDF
TokAI - TikTok AI Agent : The First AI Application That Analyzes 10,000+ Vira...
PDF
Profit Center Accounting in SAP S/4HANA, S4F28 Col11
PDF
Per capita expenditure prediction using model stacking based on satellite ima...
DOCX
The AUB Centre for AI in Media Proposal.docx
PDF
Electronic commerce courselecture one. Pdf
PDF
Mobile App Security Testing_ A Comprehensive Guide.pdf
PDF
Advanced methodologies resolving dimensionality complications for autism neur...
PDF
Approach and Philosophy of On baking technology
PPTX
Big Data Technologies - Introduction.pptx
PDF
Empathic Computing: Creating Shared Understanding
PDF
Build a system with the filesystem maintained by OSTree @ COSCUP 2025
PDF
Encapsulation_ Review paper, used for researhc scholars
PDF
Peak of Data & AI Encore- AI for Metadata and Smarter Workflows
PDF
NewMind AI Weekly Chronicles - August'25 Week I
PPTX
20250228 LYD VKU AI Blended-Learning.pptx
PPTX
Spectroscopy.pptx food analysis technology
PDF
Agricultural_Statistics_at_a_Glance_2022_0.pdf
Programs and apps: productivity, graphics, security and other tools
Blue Purple Modern Animated Computer Science Presentation.pdf.pdf
MIND Revenue Release Quarter 2 2025 Press Release
TokAI - TikTok AI Agent : The First AI Application That Analyzes 10,000+ Vira...
Profit Center Accounting in SAP S/4HANA, S4F28 Col11
Per capita expenditure prediction using model stacking based on satellite ima...
The AUB Centre for AI in Media Proposal.docx
Electronic commerce courselecture one. Pdf
Mobile App Security Testing_ A Comprehensive Guide.pdf
Advanced methodologies resolving dimensionality complications for autism neur...
Approach and Philosophy of On baking technology
Big Data Technologies - Introduction.pptx
Empathic Computing: Creating Shared Understanding
Build a system with the filesystem maintained by OSTree @ COSCUP 2025
Encapsulation_ Review paper, used for researhc scholars
Peak of Data & AI Encore- AI for Metadata and Smarter Workflows
NewMind AI Weekly Chronicles - August'25 Week I
20250228 LYD VKU AI Blended-Learning.pptx
Spectroscopy.pptx food analysis technology
Agricultural_Statistics_at_a_Glance_2022_0.pdf

Self-dependent 3D face rotational alignment using the nose region

  • 1. Self-dependent 3D Face Rotational Alignment Using the Nose Region Mehryar Emambakhsh Adrian Evans Department of Electronic and Electrical Engineering University of Bath
  • 2. Outline  Introduction to 3D alignment  SD 3D Face Rot Al  Advantages of the nose region  Algorithm steps  Preprocessing and segmentation  Nose region IFill  Energy function minimisation  Results  Conclusions and future work
  • 3. 3 Introduction to 3D alignment  3D face recognition algorithms Holistic Regional Landmark-based  3D object alignment errors Scale Translation Rotation
  • 4. 4 Introduction to 3D alignment  Alignment techniques  A rotational invariant representation  Extended Gaussian images  Spherical harmonics  Spin images  Snapshot algorithm  3D Fourier descriptors  Registration-based algorithms  Iterative closest point:  Minimum square error  Surface interpenetration measure  Self-dependent algorithms Advantages • Generalisation Disadvantages • Computational complexity • A pre-aligned reference image required • Local minima
  • 5. 5 Introduction to 3D alignment  Self-dependent algorithms Prior knowledge of the object of interest  Geometry  Curvature Much faster than the other two methods More robust for local minima No reference image required
  • 6. 6 SD 3D Face Rot Al: Advantages of the nose region  Properties of the nose region Hard to conceal without attracting suspicion Easily detected  Convexity  Nose tip, usually the closest point to the camera Small variations over different expressions
  • 7. 7 SD 3D Face Rot Al: Preprocessing and segmentation  Outliers: impulsive noise  Fuzzy C- means and median filtering  Missing data  Morphological filling and interpolation
  • 8. 8 SD 3D Face Rot Al: Nose region IFill  Filling algorithm: Applied to inverted 3D nose region
  • 9. 9 SD 3D Face Rot Al: Nose region IFill Morphological filling of the 2½D range image
  • 10. 10 SD 3D Face Rot Al: Nose region IFill Largest connected component of the filled image found:  Label with the highest frequency in the histogram.
  • 11. 11 SD 3D Face Rot Al: Energy function minimisation  Two energy (cost) functions are introduced on the found binary image domain:  EExyxy to maximise the areato maximise the area of the filled region:of the filled region:  The 3D image is rotated aroundThe 3D image is rotated around the X and Y axesthe X and Y axes  The rotation procedure stops asThe rotation procedure stops as the biggest area was foundthe biggest area was found  Simulated annealing followed bySimulated annealing followed by Levenberg-Marquart is used forLevenberg-Marquart is used for function optimisationfunction optimisation
  • 12. 12 SD 3D Face Rot Al: Energy function minimisation  EEZZ to maximise theto maximise the symmetry of the filledsymmetry of the filled region:region:  The 3D image is rotated aroundThe 3D image is rotated around the Z axisthe Z axis  The binary image is divided intoThe binary image is divided into the left and right partsthe left and right parts  The exclusive OR of isThe exclusive OR of is calculatedcalculated  The rotation procedure stopsThe rotation procedure stops when highest symmetrywhen highest symmetry obtainedobtained
  • 13. 13 SD 3D Face Rot Al: Energy function minimisation
  • 14. 14 Results  The Face Recognition Grand Challenge (FRGC) is used for evaluation. Including 557 subjects and 4947 range images.
  • 15. 15 Results  Exy minimisation The binary map in the 1st , 65th , and final iterations.
  • 17. 17 Results  Consistency evaluationConsistency evaluation  All of the images are first aligned using ourAll of the images are first aligned using our algorithmalgorithm  For each subject, each image in turn is usedFor each subject, each image in turn is used as the reference and all other images of thatas the reference and all other images of that subject aligned to it using ICP algorithmsubject aligned to it using ICP algorithm  The final rotation matrix is compared with theThe final rotation matrix is compared with the identity matrixidentity matrix
  • 18. 18 Results  Average rotation matrix for all 4947 imagesAverage rotation matrix for all 4947 images  LL22 norm of rot. matrixnorm of rot. matrix
  • 19. 19 Results  Elapsed time  Our algorithm: 2.1671 ± 0.5773s  Brute force ICP algorithm: 6.9064 ± 3.1815s
  • 20. 20 Results  Performance over different expressions
  • 21. 21 Conclusions and future work  A novel 3D face rotational alignment algorithm has been proposed  The algorithm capitalises on the nose region’s consistency, good localisation and convex features  Our approach can be used for both fine and coarse alignment purposes  Using simulated annealing as a global optimisation algorithm helps to avoid local minima & the Levenberg-Marquart algorithm performs the final tuning for B(x,y)
  • 22. 22 Conclusions and future work  Future work  The use of the final binary maps to improve for the nose segmentation  The use of the alignment algorithm for 3D face recognition will be evaluated  The final binary maps have the potential for better landmark localisation for the nose region  A more accurate computational complexity analysis is required