Research map / 3D vision / Gaussian splatting
GoMAvatar: Efficient Animatable Human Modeling from Monocular Video Using Gaussians-on-Mesh
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces GoMAvatar, which attaches local Gaussians to the faces of a skeleton-driven deformable mesh to build an animatable, real-time, memory-efficient human avatar from a single monocular video.
Problem
Creating animatable, high-quality human avatars from a single monocular video while being real-time renderable, memory-efficient, and compatible with rasterization graphics pipelines.
Approach
Gaussians-on-Mesh (GoM) attaches a Gaussian to each mesh face, defined in the triangle's local frame so it deforms with articulation; the mesh gives forward LBS articulation and geometric regularization, and color is factorized into a Gaussian-splatted pseudo albedo times a pseudo shading map predicted from the rasterized normal map.
Key results
- On ZJU-MoCap, GoMAvatar reaches 30.37 PSNR for novel view synthesis and 30.34 for novel pose synthesis, with 23.2 ms/frame (43 FPS) inference and 3.63 MB memory.
- On PeopleSnapshot, GoMAvatar achieves 30.68 PSNR / 0.9767 SSIM, versus 28.61 for InstantAvatar and 28.89 for Anim-NeRF, at 25.82 ms per frame.
- Ablation shows Gaussians-on-Mesh (30.36 PSNR) outperforms 3D Gaussians alone (30.06) and mesh alone (28.93) on novel view synthesis.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GoMAvatar instead of MonoHuman | ZJU-MoCap · novel view synthesis | PSNR: 30.26 -> 30.37 | Helped |
| use GoMAvatar instead of MonoHuman | ZJU-MoCap · novel pose synthesis | PSNR: 30.05 -> 30.34 | Helped |
| use GoMAvatar instead of MonoHuman | ZJU-MoCap · rendering | inference time (ms/frame): 5970.0 -> 23.2 | Helped |
| use GoMAvatar instead of InstantAvatar | PeopleSnapshot · novel view synthesis | PSNR: 28.61 -> 30.68 | Helped |
| remove the mesh and use 3D Gaussians only (ablation) | GoMAvatar · ZJU-MoCap novel view synthesis | PSNR: 30.36 -> 30.06 | Hurt |
| remove Gaussians and use a mesh only (ablation) | GoMAvatar · ZJU-MoCap novel view synthesis | PSNR: 30.36 -> 28.93 | Hurt |
| learn Gaussian rotation and scale in world coordinates rather than local triangle coordinates (ablation) | GoMAvatar · ZJU-MoCap novel view synthesis | Chamfer distance: 3.0728 -> 4.3941 | Hurt |
| fix local Gaussian rotation and scale (ablation) | GoMAvatar · ZJU-MoCap novel view synthesis | PSNR: 30.36 -> 30.27 | Hurt |
| remove the shading module (ablation) | GoMAvatar · ZJU-MoCap novel view synthesis | PSNR: 30.36 -> 30.13 | Hurt |
| add GoM subdivision (ablation) | GoMAvatar · ZJU-MoCap novel view synthesis | LPIPS*: 33.28 -> 32.53 | Helped |
| add GoM subdivision (ablation) | GoMAvatar · ZJU-MoCap novel view synthesis | Chamfer distance: 3.0728 -> 2.8364 | Helped |
Surprising findings
- Chamfer distance is slightly worse than HumanNeRF and MonoHuman, attributed to Gaussian thickness along the surface normal making rendered masks larger than the true silhouette.
Limitations
- Chamfer distance slightly worse than HumanNeRF and MonoHuman due to Gaussian thickness.
- Geometry pseudo ground truth is generated from NeuS rather than true scans.
- Subdivision increases inference time from 17.5 ms to 23.2 ms per frame.
Experimental setup
Datasets: ZJU-MoCap PeopleSnapshot YouTube dancing videos
Benchmarks: ZJU-MoCap PeopleSnapshot
Metrics: PSNR SSIM LPIPS Chamfer distance normal consistency inference time (ms) memory (MB)
Baselines: Neural Body HumanNeRF NeuMan MonoHuman Anim-NeRF InstantAvatar
Builds on: 3D Gaussian splatting HumanNeRF SMPL SoftRasterizer LPIPS
Closest papers by meaning
- GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians (CVPR 2024 · similarity 0.90)
- GAvatar: Animatable 3D Gaussian Avatars with Implicit Mesh Learning (CVPR 2024 · similarity 0.89)
- GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians (CVPR 2024 · similarity 0.88)
- AniGS: Animatable Gaussian Avatar from a Single Image with Inconsistent Gaussian Reconstruction (CVPR 2025 · similarity 0.87)
- MPMAvatar: Learning 3D Gaussian Avatars with Accurate and Robust Physics-Based Dynamics (NeurIPS 2025 · similarity 0.87)
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.