Exploring the Discriminative Power of SMPL-Derived Body Proportions: An Initial Investigation

Identifying individuals in unconstrained surveillance remains challenging as traditional biometrics, such as facial recognition, frequently fail due to occlusion or low resolution. This study evaluates 3D body proportions as a foundational soft biometric modality for scenarios where primary identifiers are unavailable. We propose a pipeline that extracts 23 anatomically significant joint distances from Skinned Multi-Person Linear (SMPL) models reconstructed from monocular RGB imagery. To overcome raw feature entanglement, an embedding neural network is implemented using triplet and quadruplet loss functions to map normalized proportions into a discriminative hyperspherical space. Experimental evaluation across 242 distinct identities demonstrates that the framework effectively disentangles morphological traits from estimation noise. On ground-truth 3D meshes, the system achieves approximately 95% Top-10 identification accuracy. When applied to challenging 2D-to-3D reconstructions from monocular images, the network maintains robustness, placing the correct identity within the top five candidates over 90% of the time. These results establish 3D skeletal proportions as a reliable, privacy-preserving surrogate for scalable person identification in forensic and security applications.