TL;DR
This paper presents a real-time facial age estimation system that achieves state-of-the-art accuracy using Fisher Vectors and L2-regularized linear regression. The system is efficient enough to run at 20 FPS on a single thread of a modest processor, making it suitable for practical applications.
What Problem Does This Paper Address?
Accurately estimating a person's age from their facial image has numerous applications, from age-appropriate content filtering and targeted advertising to biometric security and human-computer interaction. However, achieving this in real-time, especially on consumer-grade hardware, has been a significant challenge. Previous methods often struggled with either accuracy or speed, or required computationally expensive techniques. This paper aims to bridge this gap by proposing an efficient yet accurate system that can perform age estimation quickly.
Methods
The proposed system employs a pipeline that starts with face detection and alignment. For pre-processing, they use basic histogram equalization to improve face detection and landmark localization, especially in challenging lighting conditions. Face alignment is crucial for making the representation invariant to pose, and they use an affine transformation based on eye and mouth centers to achieve this.
The core of their approach lies in the face representation using Fisher Vectors (FV). They extract dense, multi-scale SIFT (Scale-Invariant Feature Transform) descriptors from the aligned face patches. These descriptors are then augmented with their spatial coordinates. Before computing the FV, they learn a dictionary of 64 PCA components from a large set of SIFT features. The Fisher Vector is then computed by concatenating gradients based on a Gaussian Mixture Model, effectively summarizing the distribution of local features.
For age estimation itself, they opt for L2-regularized linear regression trained using Stochastic Gradient Descent (SGD). This choice is motivated by the high dimensionality of the FV features and the desire for a computationally inexpensive prediction step, which involves a single dot product per face. This regression approach directly minimizes the error in years, avoiding the quantization issues of classification-based methods.
Datasets
The primary dataset used for training and evaluation is the MORPH-II dataset, which contains over 55,000 facial images with diverse gender and ethnicity. They also mention using a set of 5,492 images from Caucasian descent for comparison with specific prior works.
Key Insights and Ideas
- Fisher Vectors for Age Estimation: The paper demonstrates that Fisher Vectors, a powerful feature representation technique, can be effectively applied to age estimation, achieving state-of-the-art results.
- Efficiency through Linear Regression: By combining dense, high-dimensional features (Fisher Vectors) with a simple linear regressor, they achieve a remarkable balance between accuracy and speed. This avoids the computational overhead of kernel methods.
- Importance of Alignment: Robust face alignment is shown to be critical for achieving accurate landmark localization and consistent feature extraction, which directly impacts age estimation performance.
- Real-time Performance on Modest Hardware: The system's ability to run at 20 FPS on a 1.6GHz i5-2467M processor highlights its practical applicability for real-world scenarios where computational resources might be limited.
- Robustness to Demographics: The method shows good performance across different genders and ethnicities without requiring specific strategies to handle cross-racial or cross-gender influences, suggesting a generalizable approach.
Key Results
- The proposed system achieves state-of-the-art results with a Mean Absolute Error (MAE) of 3.7 years on the MORPH-II dataset.
- The system runs at 20 FPS on a single thread of a 1.6GHz i5-2467M processor.
- They demonstrate that dense SIFT sampling and a large dictionary can reduce MAE, but this comes at the cost of real-time performance. They identify a configuration that offers a good trade-off, with only a 0.1-year MAE drop.
- Comparisons with other commercial systems show comparable performance, even though those systems often use more powerful hardware and multi-threading.
Limitations and Future Work
- Trade-off between Accuracy and Speed: While the system is real-time, the authors acknowledge that denser feature sampling and larger dictionaries can further reduce the MAE but conflict with the real-time requirement.
- Future Work: Not explicitly detailed, but the paper's focus on efficiency and state-of-the-art accuracy suggests potential for further optimization and deployment in various applications. The robustness across demographics also opens doors for broader applicability.
Summary updated