Scenario-based score fusion for face recognition at a distance

Pedro Tome, Julian Fiérrez, Fernando Alonso-Fernandez, Javier Ortega-GarciaWorkshops @ CVPR 2010 · DOI

Introduction

Face recognition technology has achieved high accuracy in controlled environments, such as passport photography. However, "recognition at a distance" remains a significant challenge. As the distance between the subject and the camera increases, image resolution drops, blur increases, and environmental factors like lighting become harder to manage. This paper addresses the performance drop in these "uncontrolled" scenarios by introducing an automatic "Distance Index" to drive an adaptive fusion of different recognition algorithms.

Methodology

The researchers propose a two-part solution to improve robustness across varying distances:

The Distance Index ($D$)

To estimate acquisition distance without specialized hardware, the authors utilize the ratio of the segmented face area ($A_s$) to the total image area ($A_f$). The formula $D = -\log_{10}(A_s / A_f)$ provides a numerical value that increases as the subject moves further away. Figure 6 illustrates how this index effectively categorizes images into close, medium, and far scenarios.

Distribution of the Distance Index across different acquisition scenarios.

Adaptive Score Fusion

The system employs two distinct "matchers" with complementary strengths:

  • PCA-SVM: A holistic approach using Principal Component Analysis and Support Vector Machines. This method is precise in high-resolution, close-up conditions but sensitive to image degradation.
  • DCT-GMM: A local-feature approach using Discrete Cosine Transform and Gaussian Mixture Models. This method is more robust to low resolution and environmental noise.

Instead of a simple average (sum rule), the Distance Index is used to weight the results. As the subject moves further away (higher $D$), the system automatically increases the weight of the robust DCT-GMM matcher.

Experimental Setup

The study utilizes the NIST Multiple Biometric Grand Challenge (MBGC) v2.0 Face Stills dataset, comprising 3,842 images from 147 subjects. Images were categorized into three scenarios: "close" (shoulders visible), "medium" (upper body), and "far" (full body). The data was split into a development set (56 subjects) for tuning and an evaluation set (78 subjects) for testing.

Results and Discussion

The study confirms that PCA-SVM is a "fair-weather" algorithm—it excels in ideal conditions but degrades rapidly with distance. In contrast, DCT-GMM maintains more stable performance at a distance.

Fusion Performance

The proposed adaptive fusion approach outperformed the standard sum rule in almost all scenarios. In "medium" distance testing, the adaptive fusion provided a lower Equal Error Rate (EER) than either individual matcher. Furthermore, training models on a mixture of distances (the "mix2x" protocol) consistently yielded better results than training only on high-quality, close-up images, suggesting that exposure to "bad" data during training improves generalization.

Example of a 'far' acquisition scenario where face resolution is significantly reduced.

Conclusion

This research demonstrates that the relative size of a face within a frame is a valuable piece of metadata for intelligently blending AI models. While the current system relies on accurate face segmentation and a simple linear weighting formula, it provides a foundation for more complex, non-linear mapping functions. Future work could involve applying this scenario-based fusion logic to modern Deep Learning architectures, where different sub-networks might be specialized for varying resolutions.

Summary updated

BibTeX
@inproceedings{tome2010scenario,
  author = {Pedro Tome and Julian Fiérrez and Fernando Alonso-Fernandez and Javier Ortega-Garcia},
  title = {{Scenario-based score fusion for face recognition at a distance}},
  booktitle = {Workshops @ CVPR},
  year = {2010},
  doi = {10.1109/cvprw.2010.5543231},
}
Figure