Ordinal Regression Using Noisy Pairwise Comparisons for Body Mass Index Range Estimation

Luisa F. Polanía, Glenn Fung, Dongning Wang

TL;DR

Researchers have developed a new way to estimate Body Mass Index (BMI) categories from facial images by treating the task as a "sorting" problem rather than a simple classification one. By using Siamese neural networks to perform noisy pairwise comparisons between a test face and "anchor" faces, the system achieves higher accuracy and better rank correlation than standard deep learning models.

What Problem Does This Paper Address?

Estimating BMI from a face is notoriously difficult. Factors like age, ethnicity, and lighting can drastically change how "heavy" a face looks, and facial data is only a proxy for a measure (BMI) that is fundamentally based on the whole body.

Most existing AI approaches treat this as either a regression problem (predicting a specific number) or a multi-class classification problem (predicting a category like "Normal" or "Overweight"). The authors argue that classification is a poor fit because it assumes categories are independent. In reality, BMI categories are ordinal: "Overweight" is closer to "Obese" than it is to "Underweight." By ignoring this inherent order, standard models leave valuable information on the table.

Methods

The authors cast BMI estimation as an ordinal regression problem solved through Noisy Binary Search (NBS).

1. The Comparison Operator

Instead of asking a network "What is this person's BMI?", they ask "Is person A heavier than person B?". They use a Siamese Network—two identical twin neural networks that share weights—to extract features from two images. They then concatenate these features (and their dot product) to predict the probability that one person has a higher BMI than the other.

2. Training Modes

They tested two ways to train these comparisons:

  • Mode I: Standard binary classification (Is A > B? Yes/No).
  • Mode II: Using the Bradley-Terry model, which is typically used to rank players in sports. It models the probability of one item "beating" another based on their underlying "strengths" (in this case, their actual BMIs).

3. Noisy Binary Search (NBS)

To categorize a new face, the system doesn't compare it to every image in the database. It uses binary search algorithms (NNBS and INBS) to compare the test face against "anchors"—pools of images known to be at the boundaries of BMI categories (e.g., exactly 18.5 or 25). Because the neural network's comparisons might be wrong (noisy), the search algorithm is designed to handle errors and backtrack if necessary.

Datasets

The study utilized the inmate active population dataset from the Florida Department of Corrections.

  • Size: Nearly 100,000 images (67,756 for training, 9,045 for validation, and 20,000 for testing).
  • Scope: This is one of the largest datasets ever used for BMI estimation research.
  • Filtering: The data was filtered to subjects with BMIs between 16 and 40 to match the World Health Organization (WHO) categories.

Key Insights and Ideas

  • Relative > Absolute: It is intuitively easier for both humans and AI to determine which of two people looks heavier than it is to guess a specific BMI number from a single photo.
  • The Power of Ordinality: By using a search algorithm to "insert" a face into an ordered sequence of BMI categories, the model naturally respects the relationship between categories.
  • Bradley-Terry Advantage: Training the network to predict the probability of a "win" based on the Bradley-Terry model (Mode II) generally outperformed simple binary labels, likely because it provides a more nuanced gradient during learning.

Key Results

The proposed NBS approach outperformed standard classification and regression models across several metrics:

  • Accuracy: The best-performing model (using a VGG-based Siamese network) achieved an accuracy of 0.518, compared to 0.467 for a standard AgeNet classifier and 0.45 for traditional handcrafted geometric features.
  • Kendall Rank Correlation ($\tau$): This measures how well the predicted rankings match reality. The proposed method reached a correlation of 0.45, significantly higher than the 0.13 achieved by handcrafted features and the 0.362 of standard classification.
  • Mean Absolute Error (MAE): The NBS approach consistently resulted in lower errors, meaning that when the model was wrong, it was usually off by only one category rather than making wild miscalculations.

Limitations and Future Work

  • Demographic Bias: The dataset consists of prison inmates, which may not be representative of the general population in terms of health, nutrition, or demographics.
  • Image Quality: While the dataset is large, the authors note that distinguishing between consecutive categories (like "Normal" vs. "Overweight") remains visually challenging even for the best models.
  • Computational Cost: Performing multiple pairwise comparisons for a single search is more computationally expensive than a single "forward pass" in a standard classifier. Future work could involve optimizing the number of comparisons needed to reach a confident conclusion.

Summary updated

BibTeX
@inproceedings{polania2019ordinal,
  author = {Luisa F. Polanía and Glenn Fung and Dongning Wang},
  title = {{Ordinal Regression Using Noisy Pairwise Comparisons for Body Mass Index Range Estimation}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2019},
  doi = {10.1109/wacv.2019.00088},
}
Figure