Research map / 3D vision / Face and shape
GeoAvatar: Geometrically-Consistent Multi-Person Avatar Reconstruction from Sparse Multi-View Videos
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.
Reconstructs multi-person avatars from sparse multi-view videos using planar 2D Gaussian splats in per-person canonical spaces, with surface-ordering and monocular (Sapiens) geometry priors to prevent inter-penetration.
Problem
Independent per-person avatar reconstruction from sparse multi-view videos yields inter-penetrations and poor geometry in close multi-person interactions, while volumetric neural rendering methods are computationally expensive.
Approach
Represent each avatar with planar 2D Gaussian splats articulated by SMPL-X/LBS and auxiliary deformation, order surfaces across instances using instance masks (surface ordering loss), and supervise geometry with monocular normal and depth priors from Sapiens.
Key results
- On Hi4D (e.g., backhug02), the method reaches PSNR 31.81 and CD 0.5894 versus 29.40 PSNR and CD 0.9070 for EasyMocap and 29.90 PSNR / CD 1.7710 for multi-GART.
- Sapiens gave the best monocular normal prior among those tested on talk22 (CD 0.6480, COS 0.0142), outperforming Marigold, Metric3Dv2, and ECON.
- Training takes about 10 minutes for 3000 iterations on a single RTX A6000 GPU (the conclusion says around 5 minutes).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Multi-GART with GeoAvatar | multi-person avatar reconstruction · Hi4D Backhug02, Hug01, Talk22, Dance15, Football21 | P2S: 1.2458 -> 0.4806; 1.5944 -> 0.5034; 1.2339 -> 0.5640; 1.1377 -> 0.4769; 1.3394 -> 0.5550 | Helped |
| replace Multi-GART with GeoAvatar | multi-person avatar reconstruction · Hi4D Backhug02, Hug01, Talk22, Dance15, Football21 | PSNR: 29.9026 -> 31.8115; 29.3392 -> 31.2331; 28.8042 -> 30.1691; 25.9205 -> 27.8146; 25.7526 -> 27.3634 | Helped |
| add Gaussian normal-consistency and depth-concentration losses (ablation) | GeoAvatar · talk22 | P2S: 0.5856 -> 0.5160 | Helped |
| add monocular geometry losses (ablation) | GeoAvatar · talk22 | P2S: 0.5836 -> 0.5160 | Helped |
| add Gaussian normal-consistency and depth-concentration losses (ablation) | GeoAvatar · talk22 | CD: 0.6972 -> 0.5855 | Helped |
| add monocular geometry losses (ablation) | GeoAvatar · talk22 | CD: 0.6846 -> 0.5855 | Helped |
| add surface ordering loss (ablation) | GeoAvatar · Hi4D contact patches | mIoU Low 1%: 0.6128 -> 0.7003 | Helped |
| add surface ordering loss (ablation) | GeoAvatar · Hi4D foreground | mIoU: 0.9349 -> 0.9381 | Helped |
| use Sapiens instead of no monocular normal prior | GeoAvatar · talk22 | P2S: 0.6696 -> 0.5656 | Helped |
| use Sapiens instead of no monocular depth prior | GeoAvatar · talk22 | P2S: 0.6720 -> 0.6690 | No clear effect |
Surprising findings
- Monocular depth priors gave only marginal gains over no prior (about 0.01–0.02 dB PSNR), suggesting aligning normals matters more than depth.
Limitations
- Canonical-space deformation via deformation field alone may not capture fine dynamic details like wrinkles.
- Static surface colors and normals cause smoothing when temporal appearance changes occur.
- Reliance on monocular geometric priors inherits their biases.
Experimental setup
Datasets: Hi4D
Benchmarks: Hi4D
Metrics: PSNR SSIM LPIPS P2S Chamfer distance normal cosine distance normal L2 mIoU
Baselines: EasyMocap Multi-GART
Builds on: 2D Gaussian Splatting GART MonoSDF SMPL-X SAM2
Closest papers by meaning
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting (CVPR 2024 · similarity 0.89)
- Vid2Avatar-Pro: Authentic Avatar from Videos in the Wild via Universal Prior (CVPR 2025 · similarity 0.88)
- SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting (CVPR 2024 · similarity 0.88)
- GPAvatar: High-fidelity Head Avatars by Learning Efficient Gaussian Projections (CVPR 2025 · similarity 0.88)
- Real-time High-fidelity Gaussian Human Avatars with Position-based Interpolation of Spatially Distributed MLPs (CVPR 2025 · similarity 0.88)
- FRESA: Feedforward Reconstruction of Personalized Skinned Avatars from Few Images (CVPR 2025 · 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.