4DHuman-SKEL: Anatomically Constrained Human Model

• The paper demonstrates that 4DHuman-SKEL converts refined SMPL fits into a biomechanically accurate SKEL parameterization, offering reliable pseudo ground-truth for training.
• It details a hierarchical conversion process that optimizes SKEL parameters in phases to align skin and skeletal meshes consistently.
• Empirical results on benchmarks like 3DPW and MOYO highlight significant improvements in MPJPE and PA-MPJPE, underlining its value for anatomically constrained motion analysis.

4DHuman-SKEL is a SKEL-aligned version of the large SMPL-based 4DHuman dataset that provides pseudo ground-truth SKEL parameters and meshes by converting the refined SMPL fits in 4DHuman into the anatomically constrained SKEL space (Li et al., 25 Nov 2025). It was created to supply large-scale, high-quality SKEL supervision for SKEL-based learning, particularly because SMPL’s unconstrained ball-joint kinematics can yield anatomically implausible poses, whereas SKEL introduces anatomically accurate joints and hard joint limits but originally lacked sufficient training data (Li et al., 25 Nov 2025). In this sense, 4DHuman-SKEL functions both as a converted annotation set and as an enabling substrate for anatomically constrained human motion analysis.

1. Definition and rationale

4DHuman-SKEL was introduced in conjunction with SKEL-CF, a coarse-to-fine framework for SKEL parameter estimation (Li et al., 25 Nov 2025). Its defining property is that it re-expresses 4DHuman’s refined SMPL fits in the parameterization of SKEL, a parametric 3D human model whose pose parameters are $q \in \mathbb{R}^{46}$ and shape parameters are $\beta \in \mathbb{R}^{10}$, and which outputs both the external skin surface $M \in \mathbb{R}^{6890 \times 3}$ and an internal skeletal mesh $S_k \in \mathbb{R}^{24752 \times 3}$ (Li et al., 25 Nov 2025). Relative to SMPL’s $\theta \in \mathbb{R}^{72}$ with 3-DoF ball joints everywhere, SKEL reduces the pose dimensionality to 46 and enforces anatomically meaningful constraints, realistic limits, and a bone hierarchy consistent with human anatomy (Li et al., 25 Nov 2025).

The motivation for creating 4DHuman-SKEL is explicitly biomechanical. SMPL’s joint centers and kinematic assumptions are simplified and may not correspond to functional human joints; SKEL was developed to re-rig SMPL with a biomechanics skeleton and to provide fewer, biomechanically realistic degrees of freedom (Keller et al., 8 Sep 2025). 4DHuman-SKEL fills the resulting data gap for supervised learning by converting CameraHMR-refined 4DHuman annotations into SKEL-compatible pseudo labels (Li et al., 25 Nov 2025).

The resulting dataset is therefore not an independently captured corpus. It is a derived annotation space layered onto 4DHuman. A plausible implication is that its strengths and weaknesses are inherited jointly from the source 4DHuman fits, the CameraHMR refinement, and the SMPL-to-SKEL conversion protocol.

2. Derivation from SMPL-based 4DHuman

The source annotations come from 4DHuman, which provides pseudo-GT SMPL parameters fitted to detected 2D keypoints; CameraHMR further refines these fits through CamSMPLify and perspective modeling (Li et al., 25 Nov 2025). SKEL-CF begins from these refined SMPL meshes and applies a SKEL fitting protocol to each instance.

The conversion procedure is hierarchical. First, the SMPL mesh is reconstructed from the provided SMPL parameters. Next, SKEL parameters $(\beta, \theta, \pi)$ are initialized, while global orientation and translation are shared with SMPL and fixed during optimization; specifically, SMPL global orientation in axis-angle form is converted to SKEL’s Euler-angle convention (Li et al., 25 Nov 2025). The optimization then proceeds in phases: upper-limb refinement; full-body refinement with fixed root; and final unconstrained fine-tuning for precise mesh alignment (Li et al., 25 Nov 2025). Throughout, the optimization iteratively adjusts SKEL pose parameters under SKEL’s joint constraints to minimize discrepancy to the reconstructed SMPL mesh at both mesh level and joint level (Li et al., 25 Nov 2025).

The paper summarizes the process as:

1. $$M_{\text{smpl}} \leftarrow \text{SMPL}(\theta_{\text{smpl}}, \beta_{\text{smpl}})$$
2. Convert SMPL global orientation from axis-angle to SKEL Euler and fix global orientation and translation
3. Initialize SKEL $(\beta, \theta, \pi)$
4. Optimize over phases $$\{\text{upper\_limbs}, \text{full\_body\_fixed\_root}, \text{fine\_tune}\}$$ using $$\text{Align}(\text{SKEL}(\theta,\beta), M_{\text{smpl}})$$
5. Save $\beta \in \mathbb{R}^{10}$0, and derived joints $\beta \in \mathbb{R}^{10}$1 (Li et al., 25 Nov 2025)

Forward kinematics is central to this synchronization of skin and skeleton. In SKEL-CF, recursive global transforms are written as

$\beta \in \mathbb{R}^{10}$2

$\beta \in \mathbb{R}^{10}$3

and linear blend skinning is described conceptually by

$\beta \in \mathbb{R}^{10}$4

These relations ensure that the skeletal mesh $\beta \in \mathbb{R}^{10}$5 and skin mesh $\beta \in \mathbb{R}^{10}$6 articulate consistently during the SMPL-to-SKEL conversion (Li et al., 25 Nov 2025).

3. Data modalities and annotation schema

4DHuman-SKEL provides parameter-level and geometry-level supervision for SKEL-based learning (Li et al., 25 Nov 2025). The modalities explicitly listed are concise and technically specific.

Modality	Description
SKEL pose	$\beta \in \mathbb{R}^{10}$7 pseudo ground truth, 46-D
SKEL shape	$\beta \in \mathbb{R}^{10}$8 pseudo ground truth, 10-D
3D joints	Reconstructed $\beta \in \mathbb{R}^{10}$9
2D joints	Projected $M \in \mathbb{R}^{6890 \times 3}$0 paired with off-the-shelf 2D keypoint detections
Surface mesh	External body surface $M \in \mathbb{R}^{6890 \times 3}$1
Internal skeleton mesh	Internal skeletal mesh $M \in \mathbb{R}^{6890 \times 3}$2
Kinematic structure	FK-compatible chain consistent with SKEL hierarchy

Camera parameters occupy a distinct status. Camera extrinsics $M \in \mathbb{R}^{6890 \times 3}$3 are predicted during training, but no camera ground truth is provided in the dataset (Li et al., 25 Nov 2025). This is a consequential design decision because it separates anatomical supervision from camera supervision and forces learning of camera terms through 2D reprojection consistency alone.

The main paper does not enumerate subject counts, sequence counts, frame counts, train/validation/test layouts, or directory-level file formats for 4DHuman-SKEL (Li et al., 25 Nov 2025). It states only that 4DHuman-SKEL is constructed from the refined 4DHuman annotations of CameraHMR and provides reliable parameter-level supervision for SKEL. This absence of detailed packaging information is material: the dataset is technically defined by its converted supervisory content rather than by a fully specified standalone release format.

From a representation standpoint, 4DHuman-SKEL should be distinguished from raw SMPL corpora. SKEL explicitly models constrained human articulation with a root pelvis, spine chain, bilateral lower limbs, and bilateral upper limbs, and it outputs an internal skeletal mesh in addition to the skin (Li et al., 25 Nov 2025). This suggests that 4DHuman-SKEL is suitable not only for surface reconstruction tasks but also for analyses in which internal kinematics and anatomical plausibility matter.

4. Role in SKEL-CF training and camera-aware estimation

Within SKEL-CF, 4DHuman-SKEL is the principal training source that enables coarse-to-fine estimation of SKEL pose $M \in \mathbb{R}^{6890 \times 3}$4 and shape $M \in \mathbb{R}^{6890 \times 3}$5 (Li et al., 25 Nov 2025). The model architecture comprises a ViTPose-H encoder with 32 transformer layers, 16 heads, and hidden dimension 1280, followed by a 6-layer transformer decoder that progressively refines parameters layer by layer (Li et al., 25 Nov 2025). The encoder predicts coarse camera and SKEL parameters, and the decoder produces successive refinements.

The supervision enabled by 4DHuman-SKEL is explicit:

$M \in \mathbb{R}^{6890 \times 3}$6

$M \in \mathbb{R}^{6890 \times 3}$7

$M \in \mathbb{R}^{6890 \times 3}$8

These are combined as

$M \in \mathbb{R}^{6890 \times 3}$9

with layer-wise decoder refinement

$S_k \in \mathbb{R}^{24752 \times 3}$0

and total objective

$S_k \in \mathbb{R}^{24752 \times 3}$1

where $S_k \in \mathbb{R}^{24752 \times 3}$2 and $S_k \in \mathbb{R}^{24752 \times 3}$3 (Li et al., 25 Nov 2025).

A notable aspect of SKEL-CF is explicit camera modeling. To mitigate depth and scale ambiguities, it incorporates pinhole projection with learned intrinsics and regressed extrinsics, trained only through 2D keypoint consistency:

$S_k \in \mathbb{R}^{24752 \times 3}$4

The paper emphasizes that camera parameters are not directly supervised in 4DHuman-SKEL; instead, they are inferred from reprojection consistency (Li et al., 25 Nov 2025). This is important because it means the dataset’s core value lies in anatomical supervision, while perspective resolution remains weakly supervised.

The practical training loop therefore couples parameter supervision from 4DHuman-SKEL with differentiable forward kinematics and camera-aware reprojection. A plausible implication is that the dataset is most effective when used in systems that explicitly model camera ambiguity rather than assuming calibrated views.

5. Quantitative outcomes and empirical significance

The empirical significance of 4DHuman-SKEL is demonstrated indirectly through SKEL-CF’s performance. Trained on 4DHuman-SKEL, SKEL-CF achieves the following results:

• 3DPW: 61.5 MPJPE / 38.7 PA-MPJPE
• Human3.6M: 39.0 MPJPE / 31.2 PA-MPJPE
• MOYO: 85.0 MPJPE / 51.4 PA-MPJPE (Li et al., 25 Nov 2025)

On MOYO, the comparison to the previous SKEL-based state of the art HSMR is explicit:

• HSMR: 104.5 MPJPE / 79.6 PA-MPJPE
• SKEL-CF: 85.0 MPJPE / 51.4 PA-MPJPE (Li et al., 25 Nov 2025)

The paper reports relative improvements of $S_k \in \mathbb{R}^{24752 \times 3}$5 in MPJPE and $S_k \in \mathbb{R}^{24752 \times 3}$6 in PA-MPJPE on MOYO (Li et al., 25 Nov 2025). It further states that dense mesh errors (PVE) are competitive or superior to SMPL-based CameraHMR across multiple benchmarks and that SKEL-CF outperforms on challenging MOYO-HARD (Li et al., 25 Nov 2025).

These results are presented as evidence that large-scale parameter-level SKEL supervision is practically useful. The central claim is not merely that SKEL-CF is effective, but that 4DHuman-SKEL provides reliable $S_k \in \mathbb{R}^{24752 \times 3}$7 supervision at scale, enabling strong learning of anatomically constrained kinematics (Li et al., 25 Nov 2025). This suggests that the dataset’s principal contribution is representational: it makes direct regression into a constrained anatomical model empirically viable.

At the same time, the results should not be overgeneralized. The paper attributes success jointly to 4DHuman-SKEL and explicit camera modeling (Li et al., 25 Nov 2025). A plausible implication is that the dataset alone does not eliminate monocular ambiguities; rather, it becomes especially valuable when paired with architectures that address them.

6. Relation to broader 4D human and skeleton-driven modeling

In the literature summarized here, “4DHuman-SKEL” has a narrow and a broad meaning. In the narrow sense, it is the SKEL-aligned derivative of 4DHuman introduced for SKEL-CF (Li et al., 25 Nov 2025). In a broader conceptual sense, several later works align with a skeleton-driven 4D human workflow without defining the dataset itself.

SkeletonGaussian, for example, does not reference any dataset or method explicitly named “4DHuman-SKEL,” but it aligns conceptually with skeleton-driven 4D human modeling by representing 4D human motion via an explicit skeleton $S_k \in \mathbb{R}^{24752 \times 3}$8 and per-frame poses $S_k \in \mathbb{R}^{24752 \times 3}$9 controlling dynamic Gaussians, enabling interpretable articulated control and export to animation tools such as Blender (Wu et al., 4 Feb 2026). Motion is decomposed into skeleton-driven rigid motion via Linear Blend Skinning and HexPlane-based non-rigid refinement, thereby separating articulated structure from soft dynamics such as clothing wrinkles (Wu et al., 4 Feb 2026).

MotionDreamer likewise situates 4D human skeleton motion within a category-agnostic, video-conditioned diffusion pipeline that generates joint trajectories directly in global Cartesian space and conditions them on the target skeleton’s rest pose plus texture and semantic attributes of the mesh (Tao et al., 1 Jun 2026). Its relevance here is conceptual rather than genealogical: it addresses monocular ambiguity, heterogeneous rig structures, and exportable skeletal animation, all of which are natural downstream uses for anatomically grounded human skeleton representations (Tao et al., 1 Jun 2026).

SKEL itself supplies the anatomical substrate behind 4DHuman-SKEL. It re-rigs SMPL with a biomechanics skeleton, learns anatomical joint regressors from paired skin-skeleton data, and re-parameterizes SMPL using BSM-consistent transforms (Keller et al., 8 Sep 2025). From this perspective, 4DHuman-SKEL can be understood as a dataset-level “upgrade” of an existing human corpus into the biomechanical space introduced by SKEL, whereas methods such as SkeletonGaussian and MotionDreamer show how explicit skeletons can support editable or generative 4D pipelines.

This broader context also clarifies a common misconception. 4DHuman-SKEL is not itself a 4D generation method, a scene-aware capture benchmark, or a motion synthesis model. It is a converted supervisory resource for anatomically constrained estimation. Its relation to those neighboring problems is indirect but structurally important.

7. Limitations, failure modes, and outlook

The limitations of 4DHuman-SKEL, as stated in the SKEL-CF work, follow from its derivation pipeline rather than from a new acquisition protocol (Li et al., 25 Nov 2025). Because the labels are pseudo ground truth derived from 4DHuman and CameraHMR, they inherit biases and artifacts from noisy 2D detections, low-resolution observations, and occlusions. Although CameraHMR refines the original fits, some inaccuracies remain (Li et al., 25 Nov 2025).

A second limitation concerns camera supervision. Since no camera ground truth is provided, camera learning remains weakly supervised, so extreme perspective or limited 2D coverage can still induce depth and scale ambiguity (Li et al., 25 Nov 2025). A third limitation arises from the very constraints that motivate SKEL: constrained joint models improve realism but can underfit rare or nonstandard articulations if training coverage is limited (Li et al., 25 Nov 2025).

Failure cases described for SKEL-CF include extreme occlusions, strong ground contacts as in yoga, and unusual camera placements that degrade reprojection consistency when intrinsics are difficult to infer (Li et al., 25 Nov 2025). These are not peculiar to 4DHuman-SKEL alone, but they delimit the reliability of training and evaluation on the converted annotation space.

The future directions named in the paper are correspondingly concrete: release broader, multi-view SKEL annotations with explicit camera calibration; incorporate contact and physics priors; leverage temporal consistency across video; and expand per-joint limit learning or adaptive constraints for out-of-distribution movements (Li et al., 25 Nov 2025). This suggests a likely trajectory for the field: from converted pseudo-label corpora such as 4DHuman-SKEL toward more fully calibrated, temporally grounded, and biomechanically explicit human datasets.

In summary, 4DHuman-SKEL occupies a specific but consequential position in recent human modeling research. It is the dataset-level mechanism by which a large SMPL-based corpus is translated into an anatomically constrained SKEL parameter space, thereby enabling SKEL-CF and related biomechanics-aware estimation pipelines (Li et al., 25 Nov 2025). Its importance lies less in raw scale alone than in the type of supervision it makes available: synchronized skin, skeleton, pose, shape, and joint representations under a constrained human kinematic model.