Research map / 3D vision / Gaussian splatting
GBC-Splat: Generalizable Gaussian-Based Clothed Human Digitalization under Sparse RGB Cameras
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.
Feed-forward generalizable human digitalization from sparse binocular RGB pairs: implicit occupancy plus disparity-refined mesh, mesh-anchored 2D Gaussian surfels, and non-learned texel-based subdivision for sharper rendering.
Problem
Reconstructing complete, watertight clothed human geometry and high-quality Gaussian Splatting appearance from very sparse RGB views in a feed-forward manner, without per-subject optimization or parametric body models.
Approach
Coarse implicit occupancy field with cross-view attention, refined by stereo disparity/normal estimation and Poisson reconstruction; 2D Gaussian surfels are spawned on the mesh vertices with orientation from mesh normal and image texture gradient, then subdivided in high-frequency visible regions using clustering on projected patches without learning.
Key results
- Geometry: P2S 1.089 and Chamfer 1.218 with 8 views, outperforming GPS-Gaussian (8 views, 2.173/2.264) and 2DGS (8 views, 2.416/2.535).
- Rendering: PSNR 25.86, SSIM 0.8970, LPIPS 0.1188 with subdivision, versus 2DGS (8 views) at 25.81/0.8675/0.1719.
- Inference takes about 300 ms per frame, versus about 3 minutes for NeuS2 and 2DGS optimization.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GBC-Splat instead of GPS-Gaussian (8) | Human geometry reconstruction · Human geometry reconstruction | P2S: 2.173 -> 1.089 | Helped |
| use GBC-Splat instead of GPS-Gaussian (8) | Human geometry reconstruction · Human geometry reconstruction | Chamfer: 2.264 -> 1.218 | Helped |
| use GBC-Splat instead of 2DGS (8 Views) | Novel-view rendering · Novel-view rendering | PSNR: 25.81 -> 24.94 | Mixed |
| use GBC-Splat instead of GPS-Gaussian (8 Views) | Novel-view rendering · Novel-view rendering | LPIPS: 0.2008 -> 0.1269 | Helped |
| add Gaussian subdivision (ablation) | GBC-Splat · Novel-view rendering | PSNR: 24.94 -> 25.86 | Helped |
| add Gaussian subdivision (ablation) | GBC-Splat · Novel-view rendering | SSIM: 0.8881 -> 0.8970 | Helped |
| add Gaussian subdivision (ablation) | GBC-Splat · Novel-view rendering | LPIPS: 0.1269 -> 0.1188 | Helped |
| use 2D/3D cross-attention instead of 2D features only (ablation) | GBC-Splat geometry reconstruction · Human geometry reconstruction | P2S: 2.366 -> 1.652 | Helped |
| use 2D/3D cross-attention instead of 2D self-attention (ablation) | GBC-Splat geometry reconstruction · Human geometry reconstruction | P2S: 2.686 -> 1.652 | Helped |
| add 3D feature volume to 2D features with cross-attention (ablation) | GBC-Splat geometry reconstruction · Human geometry reconstruction | P2S: 1.698 -> 1.652 | Helped |
| add 2D/3D cross-attention (ablation) | GBC-Splat geometry reconstruction · Human geometry reconstruction | P2S: 1.103 -> 1.089 | Helped |
| add 2D/3D cross-attention (ablation) | GBC-Splat geometry reconstruction · Human geometry reconstruction | Chamfer: 1.240 -> 1.218 | Helped |
Surprising findings
- Naive 2D-feature fusion with self-attention across views degrades geometry compared with 2D features alone, attributed to severe self-occlusion under sparse inputs.
Limitations
- Frame-by-frame processing ignores temporal information, causing flickering and jittering in fast motion and complex clothing.
- Lack of topological consistency limits applications such as volumetric video compression and transmission.
Experimental setup
Datasets: Twindom THuman2.0 2K2K
Benchmarks: THuman2.0 Twindom 2K2K
Metrics: P2S Chamfer distance PSNR SSIM LPIPS
Baselines: NeuS2 2DGS DoubleField GPS-Gaussian GHG
Builds on: PIFu RAFT-Stereo 2D Gaussian Splatting Poisson Surface Reconstruction GPS-Gaussian
Closest papers by meaning
- PGC: Physics-Based Gaussian Cloth from a Single Pose (CVPR 2025 · similarity 0.91)
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.90)
- HUGS: Human Gaussian Splats (CVPR 2024 · similarity 0.88)
- ASH: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering (CVPR 2024 · similarity 0.88)
- GPS-Gaussian: Generalizable Pixel-wise 3D Gaussian Splatting for Real-time Human Novel View Synthesis (CVPR 2024 · similarity 0.88)
- GarFast: Realistic and Fast Garment Transfer with a Simplified Parser-Free Approach (AAAI 2025 · similarity 0.88)
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.