Surveillance cameras capture bodies, not faces. In the moments when a face is obscured, blurred, or captured from an unfavourable angle, traditional biometric methods fail. Can a person be reliably identified from body proportions alone — without a face, without clothing cues, without gait?
This work, presented at IIT.SRC 2026, shows that this is possible to a limited extent. The system extracts 23 anatomically significant joint distances from a 3D body model reconstructed from a standard RGB camera and uses a neural network to map them into a discriminative feature space suitable for person identification.
Method
A parametric 3D body model in SMPL format is reconstructed from monocular RGB images using Neural Localizer Fields (NLF). From this model, 22 normalised joint distances are extracted, forming a biometric signature that is independent of the subject's distance from the camera. Since the raw features exhibit high intra-class variability, an embedding neural network trained with triplet and quadruplet loss functions is proposed to map the features into a hyperspherical discriminative representation. The correct identity is then retrieved using cosine similarity.
Results
Experiments were conducted on three public datasets (HuMMan, AMASS, MVHuman) comprising 242 identities in total. On clean 3D data, the system achieved approximately 95% Top-10 accuracy even with a gallery of 160 individuals. When reconstructing from 2D images, the system correctly placed the target identity among the top five candidates in more than 90% of cases.
It is important, however, to place these results in the right context. Body proportions are among the least reliable biometric modalities — unlike fingerprints, iris patterns, or finger veins, they are considerably less unique across individuals and more prone to confusion. The method is therefore not intended as a primary identification tool, but as a fallback modality for situations where no more reliable identifier is available.
Why it matters
Despite the inherent limitations of body proportions as a biometric trait, the method offers a practically useful result: no specialised hardware required, invariant to clothing and lighting conditions, and new individuals can be added without retraining the model. Body proportions thus represent a viable fallback biometric modality for forensic and security applications — wherever primary identifiers are unavailable.
Samuel Šimún, Tomáš Goldmann, Jan Pluskal Exploring the Discriminative Power of SMPL-Derived Body Proportions: An Initial Investigation Podpořeno FIT VUT Brno, grant FIT-S-26-9011.
