Abstract
Virtual try-on and product personalization have become increasingly importantin modern online shopping, highlighting the need for accurate body measurementestimation. Although previous research has advanced in estimating 3D bodyshapes from RGB images, the task is inherently ambiguous as the observed scaleof human subjects in the images depends on two unknown factors: capturedistance and body dimensions. This ambiguity is particularly pronounced inpartial-view scenarios. To address this challenge, we propose a modular andsimple height normalization solution. This solution relocates the subjectskeleton to the desired position, thereby normalizing the scale anddisentangling the relationship between the two variables. Our experimentalresults demonstrate that integrating this technique into state-of-the-art humanmesh reconstruction models significantly enhances partial body measurementestimation. Additionally, we illustrate the applicability of this approach tomulti-view settings, showcasing its versatility.