Research map / 3D vision / Face and shape
HISR: Hybrid Implicit Surface Representation for Photorealistic 3D Human Reconstruction
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.
Proposes HISR, a hybrid implicit surface representation with two SDFs that uses surface rendering for opaque regions and volume rendering for translucent regions (e.g., hair), plus a Gaussian-mixture SDF-to-density function and a conical-frustum integrated SDF for 3D human reconstruction.
Problem
Neural implicit reconstruction and rendering of clothed humans struggle to recover heterogeneous materials (skin, hair, clothes) and fine geometric detail, while NeRF-style methods generate floating volumes and SDF-based volume methods lose fine detail.
Approach
Two SDF layers (hard opaque surface and soft outer surface) partition the scene: surface rendering on the opaque hard surface and SDF-driven volume rendering in the translucent shell, with a learnable Gaussian-mixture SDF-to-density function and an integrated SDF computed as expectation over a conical frustum.
Key results
- On WIDC novel view synthesis, HISR reaches 32.49 average PSNR vs 29.90 for NeuS and 31.85 for Gaussian Splatting.
- On WIDC geometry, HISR achieves 5.99 average single-direction CD vs 7.59 for NeuS.
- On DTU (15 scenes), average PSNR improves from 23.60 (IDR) to 25.24 and CD from 0.84 to 0.77.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HISR instead of Gaussian Splatting | 3D human reconstruction · WIDC | PSNR: 31.85 -> 32.49 | Helped |
| Use HISR instead of NeuS | 3D human reconstruction · WIDC | CD*: 7.59 -> 5.99 | Helped |
| Use HISR instead of NeuS | 3D human reconstruction · SynHuman | PSNR: 33.21 -> 35.69 | Helped |
| Use HISR instead of IDR | 3D human reconstruction · SynHuman | CD: 6.85 -> 5.66 | Helped |
| Use HISR instead of IDR | 3D object reconstruction · DTU average | PSNR: 23.60 -> 25.24 | Helped |
| Use HISR instead of IDR | 3D object reconstruction · DTU average | CD: 0.84 -> 0.77 | Helped |
| Add Gaussian Mixture density instead of Gaussian Density (ablation) | HISR · WIDC | PSNR: 29.91 -> 30.34 | Helped |
| Add Gaussian Mixture density instead of Gaussian Density (ablation) | HISR · WIDC | CD: 17.30 -> 5.99 | Helped |
| Add Gaussian Mixture density instead of Laplacian Density (ablation) | HISR · WIDC | PSNR: 29.76 -> 30.34 | Helped |
| Add Gaussian Mixture density instead of Laplacian Density (ablation) | HISR · WIDC | CD: 30.86 -> 5.99 | Helped |
| Add integrated SDF (ablation) | HISR · WIDC | PSNR: 30.28 -> 30.34 | Helped |
| Add integrated SDF (ablation) | HISR · WIDC | CD: 20.41 -> 5.99 | Helped |
Surprising findings
- Ablation with Laplace or single Gaussian density (VolSDF/NeuS-style) gives notably worse geometry, with CD rising from 5.99 to 30.86 and 17.30 respectively.
Limitations
- Quantitative evaluation on THUman not possible due to camera calibration issues, so no baseline comparison there.
- Matting masks show artifacts (e.g., foot regions) that may affect reconstruction of all methods.
- Training takes about 30 hours per instance on a single V100.
Experimental setup
Datasets: WIDC SynHuman THUman DTU
Benchmarks: WIDC SynHuman DTU
Metrics: PSNR Chamfer distance (single-direction CD*)
Baselines: NeRF IDR VolSDF NeuS Gaussian Splatting
Builds on: IDR NeuS VolSDF mip-NeRF
Closest papers by meaning
- ANIM: Accurate Neural Implicit Model for Human Reconstruction from a single RGB-D Image (CVPR 2024 · similarity 0.87)
- PSHuman: Photorealistic Single-image 3D Human Reconstruction using Cross-Scale Multiview Diffusion and Explicit Remeshing (CVPR 2025 · similarity 0.87)
- SIFU: Side-view Conditioned Implicit Function for Real-world Usable Clothed Human Reconstruction (CVPR 2024 · similarity 0.87)
- MultiGO: Towards Multi-level Geometry Learning for Monocular 3D Textured Human Reconstruction (CVPR 2025 · similarity 0.86)
- ProbeSDF: Light Field Probes For Neural Surface Reconstruction (CVPR 2025 · similarity 0.86)
- SiTH: Single-view Textured Human Reconstruction with Image-Conditioned Diffusion (CVPR 2024 · similarity 0.86)
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.