A Riemannian analysis of 3D nose shapes for partial human biometrics

Hassen Drira, Boulbaba Ben Amor, Anuj Srivastava, Mohamed DaoudiICCV 2009 · DOI

Introduction

While 3D face recognition offers significant advantages over traditional 2D imaging, it remains susceptible to two primary challenges: facial expressions and data imperfections. Expressions such as smiling or anger can drastically deform facial geometry, while laser scanners often produce "holes" in regions like the eyes and lips. This paper proposes a shift toward partial biometrics, specifically focusing on the 3D shape of the nose, which remains relatively stable regardless of expression.

The Case for Partial Biometrics

The authors argue that the nose is an ideal candidate for identification because its rigid structure is less affected by musculature changes than the mouth or eyes. By isolating this region, systems can bypass the complexities of full-face deformation.

The nose region identified as a stable feature for partial biometric analysis.

Methodology: A Riemannian Framework

To mathematically represent and compare complex 3D nasal surfaces, the researchers developed a framework based on Riemannian geometry. The process involves several sophisticated steps:

  1. Iso-curve Representation: A 3D nose scan is decomposed into a stack of "nasal curves" (level sets) representing equal distances from the nose tip.
  2. Square-Root Velocity Function (SRVF): These curves are mapped as points on a high-dimensional manifold. This allows the system to calculate "geodesics"—the shortest path or optimal deformation between two shapes.
  3. Elastic Shape Analysis: The metric used is invariant to rigid motions (rotation and translation) and re-parameterization, ensuring that different orientations or sampling rates do not interfere with identification.
  4. Hierarchical Clustering: To optimize performance, the system uses simulated annealing to organize noses into a tree structure, avoiding the need for exhaustive database searches.
3D facial scan data processed to extract nasal shape features for the Riemannian framework.

Key Insights: The "Nostril Noise" Paradox

A significant finding of the study was that more data does not always lead to better accuracy. The researchers discovered that curves passing through the nasal cavities (nostrils) introduced "undesired variability." By excluding the base and focusing only on the first seven curves (the bridge and tip), the system achieved superior results. Furthermore, the study found that the geometric mean of distances between curves outperformed the arithmetic mean, suggesting that the spatial relationship between curves is a vital biometric signal.

Results and Performance

The framework was evaluated using 2,000 3D scans from 209 individuals in the FRGC v2 database, encompassing various expressions and lighting conditions:

  • Identification: The system achieved a 76.1% rank-one recognition rate.
  • Verification: At a False Acceptance Rate (FAR) of 0.1%, the verification rate was approximately 65.34%.
  • Efficiency: The hierarchical tree structure improved retrieval speeds by 3 to 10 times compared to exhaustive searches. For instance, a search that previously took 75 seconds was reduced to just 16 seconds.

Conclusion and Future Work

While a 76% recognition rate suggests the nose is not yet a standalone replacement for full-face or iris biometrics in high-security contexts, it serves as a powerful "shortlisting" tool. The authors propose integrating this nasal analysis into a multi-modal system—combining it with data from the ears or forehead—to enhance overall accuracy. Future improvements in automated landmarking for nose-tip detection will further increase the robustness of this elastic shape analysis pipeline.

Summary updated

BibTeX
@inproceedings{drira2009riemannian,
  author = {Hassen Drira and Boulbaba Ben Amor and Anuj Srivastava and Mohamed Daoudi},
  title = {{A Riemannian analysis of 3D nose shapes for partial human biometrics}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2009},
  doi = {10.1109/iccv.2009.5459451},
}
Figure