Journee doctorant , December 12, 2012. 1 Gender and 3D Facial Symmetry: What’s the Relationship ? Xia BAIQIANG (University Lille1/LIFL) Boulbaba Ben Amor (TELECOM Lille1/LIFL) Hassen Drira (TELECOM Lille1/LIFL) Mohamed Daoudi (TELECOM Lille1/LIFL) Lahoucine Ballihi (University Lille1/LIFL) Outline 2 Introduction State-of-the-art Proposed approach Methodology Symmetry Capture Dense Scalar Field (DSF) Gender Classification Experiments Robustness to age and gender variations Robustness to expression variations Conclusions and future directions 15/03/2016 Introduction 3 Motivation to this work Why come to this idea ? Gender is essential visual attribute in human face Human faces are approximately symmetric Why use 3D face, not 2D face ? Robust to illumination and pose changes Capture more details face information 15/03/2016 State-of-the-art 4 Liu et al. used Variance Ratio (Vr) of symmetric height and orientation differences in face regions for gender classification. 111 full 3D face models were used and a result of 96.22% was achieved with a linear classifier. cooperative Based on small dataset 15/03/2016 Proposed approach 5 Reduced feature space Random Forest Adaboost SVM Training 3D scan 3D scan preprocessing Symmetry Capture (DSF) PCA-based transformation Training stage Testing stage Classification Testing 3D scan Female 15/03/2016 Symmetry Capture 6 o Represent facial surface S by a set of parameterized radial curves emanating from the nose tip. Nose tip Preprocessed face Equal angular curves extraction On the face Radial curves On the face 15/03/2016 Symmetry Capture 7 o o Corresponding symmetrical curves , . Capture symmetry by shape comparison of and . 15/03/2016 Shape Analysis of Curves 8 Represent each parameterized curve on the face, by Square-root velocity function q(t): Elastic metric is reduced to the Translations are removed Isometry under rotation & re-parameterization. metric. vs. Define the space of such functions defined as : With Norm denoted by on its tangent spaces, becomes a Riemannian manifold. Srivastava et al. TPAMI 11 15/03/2016 Geodesic Paths on Sphere 9 Geodesics in Rn are straight lines (Euclidean metric) Geodesic path connecting points p and q Derivative and module Geodesic path on Sphere Dense Scalar Field (DSF) 10 For curve and its symmetrical curve , considering the module of at each point, , located in curve with index k. With all and K considered, we build a Dense Scalar Field (DSF) for each face. 15/03/2016 Gender classification 11 High dimensional feature space 200 curves/face 100 points/curve PCA-based dimensionality reduction for SVFs Reduced subspace Machine learning Algorithm Random Forest Adaboost SVM 15/03/2016 Experiments 12 Evaluation protocol FRGC-2.0 database (UND) 466 earliest scans/4007 scans 10-fold cross validation (person-independent) 15/03/2016 Experiments 13 FRGC-2.0 database (UND) --Gender: 1848/203 females, 2159/265 males --Age : 18 to 70 (92.5% in 18-30) --Ethnicity : White 2554/319 Asian 1121/99 Other 332/48 --Expression : ~60% scans neutral --Pose : All scans in FRGC-2.0 are near-frontal. 15/03/2016 Experiments 14 (A) Robustness to age and ethnicity variations-466 scans ◦ ◦ ◦ ◦ Comparable with different classifiers Robust to number of Feature vectors Achieve 90.99% with Random forest Random Forest more effective Gender relates with face symmetry tightly Effectiveness & Robustness of approach 15/03/2016 Experiments 15 (A) Robustness to age and ethnicity variations-466 scans Observations: • Symmetrical deformation on both sides • Low deformations near symmetry plane/ high deformations faraway • female deformation changes smoother than male 15/03/2016 Experiments 16 (B) Robustness to expression variations-4007 scans ◦ Robust to number of Feature vectors ◦ Achieve 88.12% with Random forest Gender relates with face symmetry tightly Effectiveness & Robustness of Our approach 15/03/2016 Experiments 17 (B) Robustness to expression variations-4007 scans Similar observations: • Symmetrical deformation on both sides • Low deformations near symmetry plane/ high deformations faraway • female deformation changes smoother than male 15/03/2016 Comparison with state-of-the-art 18 15/03/2016 Comparison with state-of-the-art 19 General Comparison [8], [7] , [5] based on small Dataset [8], [7], [6], [5] require manual landmarking [9], [8], [7], [5] not 10-fold cross-validation Comparison with Nearest works Work1 achieves higher result than [20] with 466 scans Work2 uses whole FRGC-2.0 other than 3676 scans in [15] Weak point Dependence to upright-frontal scans. 15/03/2016 Summary and conclusions 20 Propose a fully-automatic bilateral symmetry-based 3D face gender classification approach using DSF, which is also robust to age, ethnicity and expression variations. Achieve comparable results with state-of-art, 90.99% ± 5.99 for 466 earliest scans 88.12% ± 5.53 on whole FRGC-2.0. Demonstrate that significant relationship exists between gender and 3D facial Asymmetry. 15/03/2016 Future directions 21 Deal with pose variation and incomplete data Compute more descriptors Fusion methods Gradient Spatial Symmetry Combining texture and shape, and 2D/3D methods collaboration with Chinese partners. Using symmetry-based approach for other related areas . (Age estimation result : 74% , 466 scans) 15/03/2016 Publication 22 Xia BAIQIANG ,Boulbaba Ben Amor ,Hassen,Mohamed Daoudi ,Lahoucine Ballihi, “Gender and 3D Facial Symmetry What’s the Relationship?” ,The 10th IEEE Conference on Automatic Face and Gesture Recognition, 2013. End 23 15/03/2016