Introduction
Estimating Body Mass Index (BMI) from facial images is a challenging computer vision task due to the high variability in facial appearance caused by age, ethnicity, and lighting. Unlike standard classification tasks, BMI categories (Underweight, Normal, Overweight, Obese) possess an inherent ordinal relationship. Traditional methods often treat these categories as independent classes or attempt direct numerical regression, both of which may overlook the structural information provided by the ordering of categories. This paper proposes a novel approach that frames BMI estimation as an ordinal regression problem solved through noisy pairwise comparisons.
Methodology: Pairwise Comparisons and Noisy Binary Search
The core of the proposed system is a Siamese Network designed to perform binary comparisons between two facial images. Instead of predicting an absolute BMI value, the network determines which of two individuals appears "heavier."
The Comparison Operator
The architecture utilizes two identical truncated CNNs with shared weights to extract feature vectors from input images. These vectors, along with their dot product, are concatenated and passed through fully connected layers to output a probability. The authors explored two training modes:
- Mode I: Standard binary classification (Is A > B?).
- Mode II: A Bradley-Terry model approach, which treats comparisons as probabilistic "wins" based on underlying latent strengths (actual BMIs).
Noisy Binary Search (NBS)
To categorize a test image, the system employs Noisy Binary Search. The test image is compared against "anchor" images representing category boundaries. Because neural network outputs can be noisy, the search algorithm (NNBS/INBS) is designed to handle potential errors by allowing backtracking, ensuring a robust final classification into the correct BMI range.
Dataset and Experimental Setup
The study leverages a substantial dataset from the Florida Department of Corrections, containing nearly 100,000 facial images. This dataset provides a diverse range of subjects, though it is specifically constrained to an inmate population.
The data was filtered to include BMIs between 16 and 40, aligning with World Health Organization (WHO) standards. The large scale of this dataset (67,756 training images) allows for more reliable deep learning compared to smaller, publicly available facial BMI datasets.
Results and Discussion
The NBS approach demonstrated superior performance over traditional multi-class classification and regression baselines across several key metrics:
- Accuracy: The best-performing VGG-based Siamese model achieved an accuracy of 0.518, outperforming standard AgeNet classifiers (0.467) and handcrafted geometric features (0.45).
- Rank Correlation: The Kendall Rank Correlation ($\tau$) reached 0.45, a significant improvement over the 0.362 of standard classification, indicating that the model's predictions align much more closely with the true ordinal sequence.
- Error Distribution: Mean Absolute Error (MAE) was consistently lower, with the model rarely making "wild" mispredictions (e.g., confusing Underweight with Obese).
Conclusion
By treating BMI estimation as a sorting problem rather than a simple labeling task, the authors demonstrate that relative comparisons are a more effective proxy for ordinal data. While the computational cost of multiple comparisons is higher than a single forward pass, the gains in accuracy and rank consistency suggest that noisy pairwise comparisons are a promising direction for estimating biological traits from visual data.
Summary updated