Age-Inclusive 3D Human Mesh Recovery for Action-Preserving Data Anonymization

TL;DR

Age-inclusive 3D Human Mesh Recovery (AionHMR) addresses the pediatric generalization gap in single-image 3D body reconstruction by coupling an optimization-based SMPL-A fitting stage (AionHMR-a) to generate high-quality pseudo-ground-truth with a fast transformer-based regressor (AionHMR-b) trained on this data. The framework extends 3D human mesh recovery across infants, children, and adults and demonstrates improved shape accuracy for non-adults while maintaining adult performance. It also enables privacy-preserving data sharing by releasing 3D reconstructions (3D-BabyRobot) instead of raw videos, validated through action-preservation analyses with LVLMs/LLMs. The work introduces substantial datasets and benchmarks for age-diverse motion analysis and sets a foundation for ethical, inclusive 3D human modeling and anonymized data release.

Abstract

While three-dimensional (3D) shape and pose estimation is a highly researched area that has yielded significant advances, the resulting methods, despite performing well for the adult population, generally fail to generalize effectively to children and infants. This paper addresses this challenge by introducing AionHMR, a comprehensive framework designed to bridge this domain gap. We propose an optimization-based method that extends a top-performing model by incorporating the SMPL-A body model, enabling the concurrent and accurate modeling of adults, children, and infants. Leveraging this approach, we generated pseudo-ground-truth annotations for publicly available child and infant image databases. Using these new training data, we then developed and trained a specialized transformer-based deep learning model capable of real-time 3D age-inclusive human reconstruction. Extensive experiments demonstrate that our methods significantly improve shape and pose estimation for children and infants without compromising accuracy on adults. Importantly, our reconstructed meshes serve as privacy-preserving substitutes for raw images, retaining essential action, pose, and geometry information while enabling anonymized datasets release. As a demonstration, we introduce the 3D-BabyRobot dataset, a collection of action-preserving 3D reconstructions of children interacting with robots. This work bridges a crucial domain gap and establishes a foundation for inclusive, privacy-aware, and age-diverse 3D human modeling.

Figures (7)

• Figure 1: Age-Inclusive Human Mesh Recovery with AionHMR. We present AionHMR, a novel framework for age-inclusive human mesh recovery from single RGB images. The framework consists of two stages: AionHMR-a, an optimization-based method that estimates SMPL-A shape and pose parameters along with camera parameters; and AionHMR-b, a transformer-based, real-time HMR network trained using pseudo-ground-truth annotations generated by AionHMR-a. We display AionHMR-a fittings (left, blue) and AionHMR-b results (right, gray). The bottom right shows 3D-BabyRobot database samples of child-robot interaction reconstructions.
• Figure 2: AionHMR pipeline. AionHMR-a is an optimization-based technique for the estimation of SMPL-A shape and pose parameters, and camera parameters based on a 2D reprojection loss of the keypoints. We use AionHMR-a to annotate images and create datasets. Using these datasets, we train a Transformer-based model, AionHMR-b, that regresses in real-time the same SMPL-A and camera parameters.
• Figure 3: Visual comparison of 3D human mesh estimation results. AionHMR, clearly, generates the meshes with the most accurate children's shape, while the pose and the adult reconstruction are also accurate.
• Figure 4: $\textbf{3D-BabyRobot Dataset}$. We release the 3D-BabyRobot dataset which contains 3D reconstructions of children interacting with robots. These reconstructions preserve the action and the behavior of the children and enable the release of sensitive data via anonymization.
• Figure 5: AionHMR and Baselines Comparisons Examples. From left to right: Original Image, AionHMR, HMR2.0 and BEV. Compared to HMR2.0 and BEV, AionHMR provides the most accurate estimation of the child's shape, while effectively estimating the pose.