Hybrid clustering based 3 d face modeling upon non-perfect orthogonality

Post on 26-May-2015

369 views 0 download

Tags:

description

2010 International Conference on Computer Information Systems and Industrial Management Applications (CISIM)

Transcript of Hybrid clustering based 3 d face modeling upon non-perfect orthogonality

Hybrid Clustering-based 3D Face Modeling upon Non-Perfect Orthogonality ofFrontal and Profile Views

2010 International Conference on Computer Information Systems and Industrial Management Applications (CISIM)

Alireza Ghahari, Mohsen Mosleh

Outline

Introduction Method:

Feature Global MatchingFeature Extraction In The Frontal View3D Computations In The Profile ViewsGeneric Model Modification And Texture Mapping

Introduction

Feature Global Matching

1. MIT-CBCL database (frontal view and side view)2. Histogram equalization3. A knowledge-based approach to develop a

method for micro features assembly localization.

4. Sequential Cluster Detection Algorithm

Preprocessing

Feature Clusters

ExtractionSCDAInput

Sequential Cluster Detection Algorithm

Feature Extraction In The Frontal View• Extract allotted number of feature points, those which are the most characteristic points used to represent a face.• Shows 60 marked feature points in the frontal view • Shows 31 corresponding feature points in the visible profile view.

• Using an edge detector operator, We provide edge linking for the Canny-processed images as a by-product of one double thresholding scheme to reduce the effect of false positive errors.

Double thresholding

3D Computations In The Profile Views

Make correspondences between 2D feature points, (X, Y), on frontal view and ones on the visible profile view, (Z, Y).

We then make use of symmetry between facial features which scatter at each side of the face and the resulted coordinates from the frontal view and the visible profile view to estimate the Z coordinate of features in the hidden profile view.

- A visible Profile View

• Using a 3*3 intensity correlation kernel, the algorithm self organizes the margin ‘d’ in order to force the solution Z coordinate to lie toward the middle of the d=0 solution region in hopes of improving generalization of resulting depth estimator.

• In order to minimize the errors that may be introduced by the orthogonality constraint, we utilize an algorithm on detected 2D features aiming to decrease the dependency of orthogonality of face images.

Generic Model Modification And Texture Mapping• Using NURBS curve segments for vertices definition and Delaunay triangulation solution as the connection principle, a modified version of the wellknown Candide model with 140 vertices and 264 surfaces has been designed.

※(NURBS, nonuniform rational B-splines)

• In order to reduce the effects of non-perfect orthogonality, a two step process should be considered to provide properly smooth texture image:

• 1) Image deformation: Defining two feature lines(right and left lines) on the frontal image, the profile image is deformed to be connected to the frontal image. The profile image is deformed with translation to match to the right feature line, while flipping it across to the vertical axis matches the deformed profile image to be connected to the left feature line.• 2) Image mosaic: Mosaicing of boundaries between the three combined images has been done to resist non-coherent luminance. The mosaicing is obtained using three-level wavelet packet decomposition.

End