AttriBE: Quantifying Attribute Expressivity in Body Embeddings for Recognition and Identification

A team led by Basudha Pal from Johns Hopkins University, in collaboration with Intel Labs, has developed AttriBE—a framework that quantifies attribute expressivity in body embeddings used for person re-identification (ReID). By extending the concept of expressivity via mutual information between learned features and specific attributes, the researchers used a secondary neural network to measure how strongly attributes like gender, pose (pitch/yaw), and BMI are encoded. They applied AttriBE to three transformer-based ReID models on a large-scale visible-spectrum dataset. Results show that BMI consistently exhibits the highest expressivity in deeper layers, with the final representation hierarchy ranking as BMI > Pitch > Gender > Yaw. Pose peaks in intermediate layers while BMI strengthens with depth, and attribute encoding evolves across training epochs.

The study further extends this analysis to cross-spectral person identification across short-wave, medium-wave, and long-wave infrared modalities. In this setting, pitch becomes comparable to BMI, and attribute encoding trends increase monotonically across depth, indicating greater reliance on structural cues when bridging modality gaps. This work exposes a critical implicit bias in modern ReID systems—morphometric attributes like BMI are persistently embedded, while pose plays a larger role under cross-spectral conditions. The findings have direct implications for fairness, accountability, and generalization in surveillance, security, and biometric identification applications.