A Quantitative Method for Shoulder Presentation Evaluation in Biometric Identity Documents

TL;DR

This paper introduces the Shoulder Presentation Evaluation (SPE), a quantitative method to assess shoulder square-ness in biometric identity documents using only two 3D shoulder landmarks from MediaPipe Pose. SPE yields two interpretable scores for horizontal alignment ($s_{yaw}$) and shoulder tilt ($s_{roll}$), which combine into an overall shoulder-squareness metric, and runs in real time (>100 fps). On a 121-portrait dataset, SPE shows strong agreement with human yaw judgments ($r \approx 0.80$) and supports an adapted Error-versus-Discard (EDC) analysis to quantify its utility for compliance filtering. The work demonstrates SPE as a practical tool for automated shoulder presentation checks, while noting limitations and outlining future work to integrate with head-pose assessment and real FNMR-based evaluation.

Abstract

International standards for biometric identity documents mandate strict compliance with pose requirements, including the square presentation of a subject's shoulders. However, the literature on automated quality assessment offers few quantitative methods for evaluating this specific attribute. This paper proposes a Shoulder Presentation Evaluation (SPE) algorithm to address this gap. The method quantifies shoulder yaw and roll using only the 3D coordinates of two shoulder landmarks provided by common pose estimation frameworks. The algorithm was evaluated on a dataset of 121 portrait images. The resulting SPE scores demonstrated a strong Pearson correlation (r approx. 0.80) with human-assigned labels. An analysis of the metric's filtering performance, using an adapted Error-versus-Discard methodology, confirmed its utility in identifying non-compliant samples. The proposed algorithm is a viable lightweight tool for automated compliance checking in enrolment systems.

Figures (4)

• Figure 1: Example results of the Shoulder-Presentation Evaluation (SPE) algorithm. Each of the seven panels shows a test image with MediaPipe Pose landmarks overlaid. The evaluation panel below each image compares the algorithm's computed horizontalScore and shoulderTiltScore against the human-annotated manualLabel (ground truth). Debug information, including the calculated yaw angle and the absError between the horizontalScore and manualLabel, illustrates the model's internal calculations. The examples range from fully compliant (left, score = 1.0) to moderately rotated torsos, demonstrating a strong correlation between the algorithm's scores and perceptual judgments.
• Figure 2: Scatter plot of the SPE algorithm’s output score vs. the manual label for shoulder alignment on 121 test images. The dashed line is the identity line (perfect prediction). The Pearson correlation is $\approx$0.80 and mean absolute error is 0.10, indicating a strong agreement between the automatic metric and human judgement.
• Figure 3: Distribution of absolute errors ($|$manual – algorithm score$|$) across the test set. The histogram shows that in the majority of cases, the error is very low. The red dashed line highlights an error threshold of 0.2.
• Figure 4: Adapted Error-versus-Discard Characteristic (EDC) for the shoulder-alignment quality score. The y-axis shows the classification false-negative rate (FNR) relative to human labels, while the x-axis is the fraction of lowest-scored images discarded. This adapted EDC curve evaluates the metric's ability to filter non-compliant samples based on human judgment, rather than its effect on a specific biometric recognizer. The blue curve is the empirical SPE result; the orange dashed line is an oracle that would remove every misclassified image first. Discarding the worst 40 % of images lowers the FNR from 19.5 % to 4.3 %, illustrating the metric’s utility for compliance filtering.