🗺️ ATLAS: An Expressive and Controllable Mesh Model
Parametric body models offer expressive 3D representation of humans across a wide range of poses, shapes, and facial expressions, typically derived by learning a basis over registered 3D meshes. However, existing human mesh modeling approaches struggle to capture detailed variations across diverse body poses and shapes, largely due to limited training data diversity and restrictive modeling assumptions. Moreover, the common paradigm first optimizes the external body surface using a linear basis, then regresses internal skeletal joints from surface vertices. This approach introduces problematic dependencies between internal skeleton and outer soft tissue, limiting direct control over body height and bone lengths. To address these issues, we present ATLAS, a high-fidelity body model learned from 600k high-resolution scans captured using 240 synchronized cameras. Unlike previous methods, we explicitly decouple the shape and skeleton bases by grounding our mesh representation in the human skeleton. This decoupling enables enhanced shape expressivity, fine-grained customization of body attributes, and keypoint fitting independent of external soft-tissue characteristics. ATLAS outperforms existing methods by fitting unseen subjects in diverse poses more accurately, and quantitative evaluations show that our non-linear pose correctives more effectively capture complex poses compared to linear models.
ATLAS reconstructs high-fidelity 3D humans across a wide range of poses, capturing both skeletal structure and surface details.
ATLAS separates the internal skeleton from the external surface, removing spurious vertex-joint correlations and enabling independent control of the shape and skeleton.
ATLAS first aligns the skeleton to keypoints, then adapts body shape to silhouettes. This produces accurate and realistic 3D meshes under diverse real-world conditions.
Core Idea: The surface vertices of the human body (e.g. soft tissue) should be independent of the internal skeleton (e.g. bone lengths).
Prior work entangle the soft surface and the internal bones: they first personalize the surface vertices, then derive internal joints as a weighted sum of the surface. This vertex-centric framework introduces undesirable correlations between the external surface and internal skeleton, and it inhibits precise, decoupled control of soft surface variation and internal skeletal attributes.
Comparison of Skeleton Symmetries. SMPL-X (left) derives joints from surface vertices, producing an asymmetric skeleton, whereas ATLAS (right) maintains a symmetric and consistent skeleton.
Controllable Skeletal Attributes. ATLAS enables easy control of skeletal parameters such as spine length, shoulder width, and hand size.
Instead, ATLAS defines an explicit default skeleton that all subjects' surface vertices are aligned to. All surface variation aligned to this default skeleton, and bone length variation is applied to the skeleton during the Linear Blend Skinning posing process. This separation enables easy control of the surface and skeleton.
Character Customization Demo. Starting with a detailed ATLAS mesh of a subject, we sequentially increase the shoulder width, change the arm length, and adjust the body weight. The resulting mesh is highly realistic, and it maintains details of the original subject shape and skeleton while naturally incorporating the added customizations.
Mesh models come with pose correctives. These correctives are pose-dependent vertex offsets applied to the mesh prior to Linear Blend Skinning (LBS) to address joint collapse and other LBS posing artefacts.
Prior work use linear correctives (insufficiently expressive) or dense, non-linear MLPs (undesirable full-body correlations). ATLAS develops sparse, non-linear pose correctives, enforcing sparse, local joint-vertex correlations while allowing non-linear, expressive entanglement between neighboring kinematic joints.
Sparse Pose Correctives. The first row displays pose correctives from SMPL-X. The second row shows the inverse geodesic initialization for our pose corrective activations, and the third row demonstrates their sparsity after convergence.
Linear vs non-linear pose correctives. For both linear and non-linear pose correctives (PCB), we show the predicted vertex offsets in the rest pose, the posed mesh, and the fitting error heatmap. Sparse, non-linear PCBs achieve lower error and yield more realistic meshes.
ATLAS is trained on a large-scale dataset of 600k high-resolution scans of subjects in diverse poses. Trained on a more diverse set of shapes, identities, and poses, ATLAS is a general, expressive human body model.
Sampled Visualizations of Our Multi-Pose Dataset. Scans are captured using a high-resolution scanner with 240 synchronized cameras. We then register meshes by regularizing joint locations with triangulated keypoints before performing iterative mesh fitting.
We pair ATLAS with a performant single-image fitting pipeline naturally compatible with our decoupled shape and skeleton model.
In our multi-stage optimization, skeleton and pose are fit first using keypoint and depth terms, followed by surface shape and expression refinement through mask and expression losses. This ordering ensures clean disentanglement of structure from soft tissue, enabling accurate body shape recovery and robust expression capture.
Results of fitting ATLAS to in-the-wild images. Our multi-stage body fitting procedure robustly handles clothed subjects in varying poses along with detailed facial expressions.
@inproceedings{park2025atlas,
title = {ATLAS: Decoupling Skeletal and Shape Parameters for Expressive Parametric Human Modeling},
author = {Park, Jinhyung and Romero, Javier and Saito, Shunsuke and Prada, Fabian and Shiratori, Takaaki and Xu, Yichen and Bogo, Federica and Yu, Shoou-I and Kitani, Kris and Khirodkar, Rawal},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
year = {2025}
}