Research map / 3D vision / Face and shape
Fine Structure-Aware Sampling: A New Sampling Training Scheme for Pixel-Aligned Implicit Models in Single-View 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 Fine Structure-aware Sampling (FSS), a sampling training scheme for pixel-aligned implicit models that captures thin body features, plus normal-of-sample-point supervision (NSP) and a mesh thickness loss (MTL) for single-view clothed human reconstruction.
Problem
Pixel-aligned implicit models for single-view clothed human reconstruction miss thin surfaces (ears, fingers) under existing sampling schemes (spatial sampling, DOS) or produce noisy wavy artefacts and implausible mesh thickness.
Approach
A sampling scheme with twinned sample points whose midpoint pinpoints surfaces, proximity-adaptive displacement, anchor and counter sample points, and SMPL-X-guided sampling density; additionally uses sample-point normals as supervision on the feature volume and a thickness loss enabled by a Tri-PIFu stack that uses trilinear interpolation and sums the volume along depth.
Key results
- FSS (Ours) achieves CD 4.833e-4 and P2S 3.854e-4 on THuman2.0, beating PIFu (5.314, 4.933) and IntegratedPIFu (5.172, 4.276 on P2S? no, P2S 4.276 for IntegratedPIFu so FSS wins on P2S too)
- HRI + FSS reports the best Normal error on both datasets (7.615e-2 on THuman2.0, 5.715e-2 on BUFF) among compared methods
- Adding NSP to PIFu reduces THuman2.0 CD from 5.314 to 4.991 and BUFF CD from 2.089 to 2.041; adding MTL reduces THuman2.0 CD to 5.131
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace spatial sampling with FSS (ablation) | PIFu · THuman2.0 | CD: 5.314 -> 5.004 | Helped |
| Replace DOS with FSS (ablation) | PIFu · THuman2.0 | CD: 5.943 -> 5.004 | Helped |
| Add NSP (ablation) | PIFu · THuman2.0 | CD: 5.314 -> 4.991 | Helped |
| Add MTL (ablation) | PIFu · THuman2.0 | P2S: 4.933 -> 3.927 | Helped |
| Remove twinned sample points (ablation) | FSS · THuman2.0 | Normal: 8.052 -> 8.728 | Hurt |
| Remove proximity-adaptive displacement (ablation) | FSS · THuman2.0 | CD: 5.004 -> 5.399 | Hurt |
| Remove anchor sample points (ablation) | FSS · THuman2.0 | P2S: 3.965 -> 4.021 | Hurt |
| Remove counter sample points (ablation) | FSS · THuman2.0 | Normal: 8.052 -> 8.854 | Hurt |
| Remove SMPL-X-guided sampling (ablation) | FSS · THuman2.0 | CD: 5.004 -> 5.135 | Hurt |
| Add HRI | FSS · THuman2.0 Test Set | Normal: 7.800 -> 7.615 | Helped |
Surprising findings
- PIFu + DOS performs worse than plain PIFu with spatial sampling on THuman2.0 CD and P2S (5.943 vs 5.314 CD)
- Removing the twinned query points ablation slightly improves P2S (3.902 vs 3.965), suggesting some features trade off against others
Limitations
- FSS-only (no HRI) model does not beat PIFuHD and IntegratedPIFu on the Normal metric on THuman2.0
- Smplx-guided sampling requires ground-truth SMPL-X meshes during training
- Evaluation is limited to THuman2.0 training and BUFF test sets with a single 80-20 split
Experimental setup
Datasets: THuman2.0 BUFF
Benchmarks: THuman2.0 BUFF
Metrics: Chamfer distance Point-to-Surface Normal reprojection error
Baselines: PIFu PIFu + DOS S-PIFu ICON PIFuHD IntegratedPIFu
Builds on: PIFu PIFuHD IntegratedPIFu pix2pixHD Stacked Hourglass
Closest papers by meaning
- SIFU: Side-view Conditioned Implicit Function for Real-world Usable Clothed Human Reconstruction (CVPR 2024 · similarity 0.86)
- Diffusion-FOF: Single-View Clothed Human Reconstruction via Diffusion-Based Fourier Occupancy Field (CVPR 2024 · similarity 0.86)
- PSHuman: Photorealistic Single-image 3D Human Reconstruction using Cross-Scale Multiview Diffusion and Explicit Remeshing (CVPR 2025 · similarity 0.86)
- ODHSR: Online Dense 3D Reconstruction of Humans and Scenes from Monocular Videos (CVPR 2025 · similarity 0.85)
- BLADE: Single-view Body Mesh Estimation through Accurate Depth Estimation (CVPR 2025 · similarity 0.85)
- MultiGO: Towards Multi-level Geometry Learning for Monocular 3D Textured Human Reconstruction (CVPR 2025 · similarity 0.85)
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.