Distinguishing Facial Features for Ethnicity-Based 3D Face Recognition

Stefano Berretti, Alberto Del Bimbo, Pietro PalaACM TIST 2012 · DOI

TL;DR

This paper introduces a novel approach to 3D face recognition that leverages feature selection to identify which facial regions are most discriminative for different ethnic groups. By analyzing the relevance of specific facial features across ethnicities and under varying expressions, the authors demonstrate that tailoring recognition models to ethnic groups can significantly improve accuracy.

What Problem Does This Paper Address?

Traditional 3D face recognition methods often treat all facial features equally or use global descriptors. While local feature-based approaches offer more robustness to variations like expressions and pose, they haven't fully explored the differential importance of various facial regions, especially concerning ethnic variations. This paper addresses the gap in understanding how specific facial features contribute to recognition accuracy across different ethnic groups and how these contributions change with facial expressions. The goal is to move beyond generic face recognition models towards more accurate, ethnicity-aware systems.

Methods

The core of this work lies in combining a 3D face representation method with a feature selection technique.

  • 3D Face Representation: The authors utilize "iso-geodesic stripes" (IGS) to represent 3D face scans. These stripes are generated by measuring geodesic distances from a reference point (nose tip) on the face surface. The face is then partitioned into three regions (upper-left, upper-right, and lower) to better handle asymmetries and expression variations.
  • Feature Description: The structural information of these IGS is captured using "3D Weighted Walkthroughs" (3DWWs), which describe the spatial relationships between points in different stripes. These 3DWWs are then used to label nodes and edges in a graph representation of the face.
  • Feature Selection: The Minimal-Redundancy Maximal-Relevance (mRMR) algorithm is employed. This filter-based method aims to select features (in this case, distances between pairs of IGS) that are highly relevant to the classification task (distinguishing between same and different individuals) while being minimally redundant with each other. This helps in identifying the most informative facial regions.
  • Face Matching: A similarity measure is defined for comparing the graph representations of two faces, combining intrastripe and interstripe 3DWW similarities.

Datasets

The experiments were conducted using the Face Recognition Grand Challenge (FRGC) dataset.

  • FRGC v1.0: Used for training and evaluating the feature selection process. It contains 943 scans from 275 individuals with neutral facial expressions.
  • FRGC v2.0: Used as the test set. It is a larger dataset with 4,007 scans from 466 individuals, including various facial expressions.

The paper also describes partitioning these datasets by ethnicity (Asian and Caucasian) to analyze ethnic-specific feature relevance.

Key Insights and Ideas

  • Differential Feature Relevance: Not all facial regions are equally important for 3D face recognition. The mRMR feature selection method effectively quantifies this relevance.
  • Ethnicity-Specific Features: The most discriminative facial features (stripe pairs) vary significantly across different ethnic groups. This suggests that a one-size-fits-all approach to 3D face recognition is suboptimal.
  • Expression Robustness: While facial expressions do cause deformations, the iso-geodesic stripes approach, especially with face partitioning, shows a degree of robustness. The feature selection process helps identify features that are less affected by expressions within specific ethnic groups.
  • Potential for Improved Accuracy: By identifying and utilizing ethnicity-specific features, it's possible to train more accurate face recognition models tailored to particular ethnic groups.

Key Results

  • The mRMR analysis identified specific pairs of iso-geodesic stripes as being most relevant for distinguishing faces.
  • The top-ranked relevant stripe pairs differed considerably between Asian and Caucasian datasets. For instance, the area around the nose (stripes 4 and 5) was found to be a discriminating trait for Caucasians but less so for Asians.
  • When tested on the FRGC v2.0 dataset, classifiers trained on homogeneous ethnic groups (Asian-v2 or Caucasian-v2) achieved higher accuracy than a classifier trained on the entire, non-homogeneous dataset (All-v2).
  • For example, at a False Acceptance Rate (FAR) of 0.001, using approximately 75 of the most relevant features provided a reasonably high True Acceptance Rate (TAR).
  • The proposed approach achieved competitive performance, with an overall TAR of 90.3% on the FRGC v2.0 dataset for the all-vs.-all experiment, close to state-of-the-art methods.
  • A significant advantage of the proposed method is its computational efficiency, with face verification being performed online in approximately 12 ms.

Limitations and Future Work

  • Ethnicity Classifier: The current work assumes a perfect ethnicity classifier. In practice, incorporating a real ethnicity classifier would likely reduce recognition accuracy by an estimated 2%.
  • Future Work: The authors propose to further investigate the use of region selection results to explicitly classify different ethnic groups and then use this information to optimize recognition results. This could lead to a two-stage recognition system: first, identify the ethnic group, and second, perform face recognition within that specific group.

Summary updated

BibTeX
@article{berretti2012distinguishing,
  author = {Stefano Berretti and Alberto Del Bimbo and Pietro Pala},
  title = {{Distinguishing Facial Features for Ethnicity-Based 3D Face Recognition}},
  journal = {ACM Trans. Intell. Syst. Technol.},
  year = {2012},
  doi = {10.1145/2168752.2168759},
}
Figure