fit parametric body model

Existing parametric human models (e.g., SMPL) employ kinematic structures inconsistent with anatomical skeletal geometry, limiting their applicability in biomechanical analysis. To address this, we propose SKEL—the first learnable parametric 3D human model explicitly incorporating biomechanically accurate skeletal anatomy. Our method re-rigs the SMPL mesh to establish a neural mapping from vertices to anatomically consistent joint locations and bone rotations. We optimize skeletal pose priors using the AMASS dataset and reparameterize mesh dynamics to enable monocular image-driven estimation of biomechanical parameters. Compared to conventional models, SKEL preserves full animatability while significantly improving joint localization accuracy and ensuring physiologically plausible constraints on degrees of freedom. Crucially, SKEL enables lossless upgrading of existing pose datasets into new biomechanics-aware benchmarks—retaining all original annotations while enriching them with anatomically grounded skeletal parameters.

Existing parametric human models (e.g., SMPL) lack biomechanical fidelity; while SKEL provides anatomically accurate skeletal structure, its parameter estimation is hindered by data scarcity, occlusion-prone multi-view ambiguity, and joint complexity. To address these challenges, we propose SKEL-CF: a Transformer-based coarse-to-fine encoder-decoder framework—where the encoder produces initial estimates of camera and SKEL parameters, and the decoder refines them hierarchically. We explicitly model camera geometry to mitigate depth and scale ambiguities. Furthermore, we introduce 4DHuman-SKEL, the first large-scale 4D dataset specifically designed for anatomical skeleton estimation. Evaluated on MOYO, SKEL-CF achieves 85.0 mm MPJPE and 51.4 mm PA-MPJPE, substantially outperforming HSMR (104.5 / 79.6), and marks the first method enabling jointly high-accuracy, anatomically plausible reconstruction of both skeletal structure and surface geometry.

Existing parametric human body models, such as SMPL and SMPL-X, are difficult to unify due to incompatibilities in mesh topology, skeletal structure, and shape parameterization. This work proposes a three-layer abstraction—mesh topology, skeleton, and pose—to construct a unified human representation that enables seamless conversion across models. By leveraging constant-time vertex mapping, closed-form joint transformation recovery, and inverse skinning for pose extraction, the method reduces the complexity of many-to-many model adaptation from O(M²) to O(M). Integrated within a fully differentiable architecture and accelerated via NVIDIA Warp on GPUs, the approach requires neither iterative optimization nor model-specific training. It supports arbitrary mixing of identity and motion from compatible models at inference time, significantly enhancing flexibility and efficiency in human reconstruction, animation, and simulation.

The SMPL model struggles to represent complex 3D geometries—such as handheld objects and loose clothing—that exhibit motion decoupling from the underlying body. Method: We propose an adaptive, modular skeletal growth framework that enables dynamic expansion of the kinematic tree. Our approach introduces a novel modular bone mechanism supporting joint-tree growth, jointly optimizing external joint poses and Linear Blend Skinning (LBS) weights in SE(3) space. It further incorporates motion-kernel-guided gradient localization, differentiable rendering, and explicit animation control. Contribution/Results: Unlike conventional unified parametric models, our framework significantly improves expressivity and controllability for decoupled components. Experiments demonstrate superior rendering fidelity and more robust animation generation across diverse challenging apparel scenarios—including voluminous garments and articulated accessories. The framework establishes a general, extensible paradigm for enhanced SMPL-based modeling, enabling scalable integration of anatomically and physically plausible non-rigid elements.

This work addresses the limitations of existing SMPL-based 3D Gaussian Splatting methods for human avatars, demonstrating that the primary bottleneck lies in the representational capacity of the body model rather than architectural complexity. The authors propose replacing SMPL with the more expressive Momentum Human Rig (MHR), integrated with poses estimated by SAM-3D-Body, to achieve high-quality reconstructions through an exceptionally streamlined pipeline that avoids learning any deformation or pose-dependent corrections. Controlled ablation studies are conducted to isolate and validate, for the first time, the critical role of body model expressiveness in reconstruction fidelity. The method achieves state-of-the-art PSNR on both PeopleSnapshot and ZJU-MoCap datasets, while attaining leading or comparable performance in LPIPS and SSIM metrics.

This study addresses the significant amplification of noise in 3D human pose estimates from monocular video when computing joint torques via inverse dynamics, particularly affecting proximal joints. The work presents the first quantitative analysis of this noise amplification effect—demonstrating an approximate 1000-fold increase—and introduces SMPL-Dynamics, a fully differentiable inverse dynamics module built upon the SMPL body model that operates without external physics simulators. By integrating low-pass filtering and differentiable pose refinement prior to differentiation, the proposed method reduces joint torque error by 93% while preserving pose estimation accuracy, thereby substantially enhancing the robustness and fidelity of torque estimation.

Traditional CT-based anatomical models are static and fail to capture morphological dynamics under postural changes, limiting their utility in radiographic imaging and preoperative planning. This work proposes a method to construct patient-specific dynamic digital twins from a single whole-body CT scan: by fitting the SMPL parametric human body model to obtain an aligned kinematic skeleton, segmented bones and organs are bound to an anatomy-aware rigging system, enabling pose re-targeting while preserving geometric consistency and generating digitally reconstructed radiographs (DRRs) in novel poses. To the best of our knowledge, this is the first approach capable of producing posture-adaptive anatomical models from static CT data. Experiments on three subjects demonstrate a skeleton fitting Chamfer distance of 15.8 ± 4.0 mm, bone coverage of 95.9 ± 1.8%, DRR similarity between original and reposed images with SSIM of 0.872 ± 0.016 and PSNR of 18.5 ± 1.4 dB, and stable organ coverage at 94.4 ± 0.4%.

This study addresses the clinical need for a convenient and low-cost method to quantify joint angles. The authors propose a novel approach that directly maps segment rotation matrices—output by arbitrary parametric human body models such as GEM-X or SAM 3D Body—to clinically relevant joint angles, without requiring inverse kinematics, musculoskeletal modeling, subject-specific calibration, height measurements, camera parameters, or individualized modeling. Relying solely on a compact calibration table, the method is computationally efficient, enables real-time processing of monocular video, and demonstrates cross-model generalizability. Evaluated on the OpenCap LabValidation dataset, it achieves a mean absolute error of 4.50 degrees, matching the performance of OpenCap Monocular.

This work addresses the gap between geometric 3D human pose estimation and the biomechanical attributes required in rehabilitation and sports science. We propose BioModule, a lightweight, pose-estimator-agnostic temporal Transformer module that can be appended to any existing 3D pose estimator to predict biomechanically meaningful quantities from standard 17-joint skeletons. To enable frame-level cross-modal supervision, we construct the first large-scale aligned dataset and systematically analyze the impact of upstream pose accuracy on downstream biomechanical prediction performance. By integrating anatomical coordinate alignment with the Human3.6M family of datasets, BioModule demonstrates consistent effectiveness across seven state-of-the-art pose estimators, enabling, for the first time, non-invasive and physically interpretable visual biomechanical analysis.