1994 view-based-and-modular-eigenspaces-for-face-recognition
The paper presents a comprehensive study of eigenface‑based face recognition applied to a large‑scale database of 7,562 images (~3,000 individuals). It first demonstrates that a 20‑eigenvector representation can achieve 95% recognition accuracy on this dataset, despite variations in expression, illumination, and pose. To address pose variability, the authors contrast two formulations: a single parametric eigenspace that encodes identity and view jointly, and a view‑based approach that maintains separate eigenspaces for each viewing angle and selects the best one via a distance‑from‑face‑space (DFFS) metric. Experiments show the view‑based method slightly outperforms the parametric one in both interpolation and extrapolation tests. The paper then introduces a modular eigenspace layer that models salient facial parts (eyes, nose, mouth) as separate eigenfeatures. An automatic DFFS‑based detector locates these parts across views, achieving up to 94% detection at 6% false‑alarm rate with ten eigenvectors. Using the modular representation, recognition rates reach 95% with eigenfeatures alone and 98% when combined with whole‑face eigenvectors, especially improving robustness to occlusions and disguises. The authors also describe Photobook, an interactive X‑window tool for browsing and similarity‑searching the database. The work concludes that eigen‑based methods scale to thousands of identities, that view‑based eigenspaces effectively handle pose, and that modular part‑based representations enhance accuracy and resilience.
Topics
Scaling Eigenfaces to a 3,000‑Person Database
Evaluation of eigenface recognition using 20 eigenvectors on a 7,562‑image dataset, achieving 95% accuracy despite expression and illumination variation.
Photobook: Interactive Similarity Search Interface
Description of the X‑window tool that lets users filter the database by metadata and retrieve the 20 most similar faces using eigenvector distances.
View‑Based vs. Parametric Eigenspace Formulations
Comparison of a single universal eigenspace with multiple view‑specific eigenspaces, including DFFS‑based view selection and performance on interpolation/extrapolation tests.
Distance‑From‑Feature‑Space (DFFS) Detector for Facial Parts
Algorithm that computes residual error of feature eigenspaces to locate eyes, nose, and mouth across poses, with ROC analysis showing 94% detection at 6% false‑alarm using ten eigenvectors.
Modular Eigenspace Representation (Eigenfeatures)
Construction of separate eigenvectors for eyes, nose, mouth and their integration with whole‑face eigenvectors, yielding up to 98% recognition and robustness to occlusions.
Empirical Performance Results and ROC Analyses
Tables and figures reporting recognition rates for whole‑face, eigenfeatures, and combined systems, as well as detection ROC curves for eyes under varying eigenvector counts.
Conclusions and Future Directions
Summary of findings—scalability, pose handling, modularity—and discussion of potential extensions such as weighted fusion and sequential coarse‑to‑fine classification.