Estimation of BMI from facial images using semantic segmentation based region-aware pooling

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Body-Mass-Index (BMI) estimation from facial images has implications for predicting health and socio-economic factors.
• Existing methods using hand-crafted features or global deep features lack generalizability and detailed local information, respectively.

Purpose of the Study:

• To enhance automatic Body-Mass-Index (BMI) estimation by leveraging detailed local facial information.
• To improve the accuracy of BMI prediction through a novel approach focusing on pooled deep features from specific facial regions.

Main Methods:

• Proposed a framework utilizing deep features pooled from distinct facial regions (e.g., eyes, nose, lips).
• Integrated face semantic segmentation for accurate pixel-level localization of facial regions.
• Employed Convolutional Neural Network (CNN) backbones such as FaceNet and VGG-face on VisualBMI, Bollywood, and VIP attributes datasets.

Main Results:

• The proposed Reg-GAP method demonstrated significant performance improvements across three public datasets.
• Achieved percentage improvements of 22.4% on VIP-attribute, 3.3% on VisualBMI, and 63.09% on the Bollywood dataset compared to recent works.
• Explicit pooling of deep features from segmented facial regions substantially boosted BMI prediction accuracy.

Conclusions:

• The proposed method effectively utilizes detailed local facial information for more accurate BMI prediction.
• Face semantic segmentation combined with regional feature pooling offers a superior approach for automatic BMI estimation.
• This technique holds potential for broader applications in predicting societal behaviors linked to BMI.