Research map / 3D vision / Face and shape
HiLo: Detailed and Robust 3D Clothed Human Reconstruction with High-and Low-Frequency Information of Parametric Models
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 HiLo, a monocular clothed-human reconstruction method that combines progressive high-frequency SDF features (for detail) with low-frequency voxelized SMPL-X features and a spatial interaction implicit function (for robustness to parametric-body noise).
Problem
Single-RGB-image 3D clothed human reconstruction struggles to be simultaneously detailed and robust to inaccurate parametric body (SMPL-X) estimates.
Approach
Uses high-frequency SDF features of the SMPL-X body, introduced progressively (coarse-to-fine weighting of sinusoidal frequency bands) to avoid large gradients, together with low-resolution voxelized SMPL-X features processed by a spatial interaction module with attention across voxels to gain robustness to body noise.
Key results
- HiLo improves Chamfer distance over state-of-the-art by 10.43% on THuman2.0 and 9.54% on CAPE, per the abstract.
- On CAPE unseen setting (trained on THuman2.0), HiLo achieves Chamfer 0.7861 vs ICON 0.8610 and D-IF 0.8332.
- Under SMPL-X noise 0.5 on CAPE, HiLo reaches Chamfer 1.3593 vs ICON 4.6716 (unseen setting).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HiLo instead of ECON | 3D clothed human reconstruction · CAPE-FP | Chamfer distance: 0.9651 -> 0.6954 | Helped |
| use HiLo instead of ECON | 3D clothed human reconstruction · CAPE-NFP | Chamfer distance: 0.9983 -> 0.7830 | Helped |
| use HiLo instead of ECON | 3D clothed human reconstruction · CAPE | Chamfer distance: 0.9872 -> 0.7430 | Helped |
| use HiLo instead of D-IF | 3D clothed human reconstruction · Thuman2.0 | Chamfer distance: 1.0305 -> 0.9230 | Helped |
| use HiLo instead of D-IF | 3D clothed human reconstruction · CAPE (train Thuman2.0, test CAPE) | Chamfer distance: 0.8332 -> 0.7861 | Helped |
| add progressive HF SDF rather than vanilla SDF (ablation) | HiLo · CAPE | Chamfer distance: 0.8118 -> 0.7430 | Helped |
| add spatial interaction implicit function (ablation) | HiLo · CAPE | Chamfer distance: 0.8555 -> 0.7430 | Helped |
| add progressive HF SDF and spatial interaction implicit function (ablation) | HiLo · CAPE | Chamfer distance: 0.9362 -> 0.7430 | Helped |
| add voxel-grid features M³ᴰv(p) (ablation) | HiLo · CAPE, SMPL-X noise=0.5 | Chamfer distance: 4.9844 -> 1.3593 | Helped |
| use progressive HF SDF instead of vanilla SDF (ablation) | HiLo · CAPE, SMPL-X noise=0.2 | Chamfer distance: 1.6909 -> 1.3427 | Helped |
| use progressive HF SDF instead of conventional high-frequency SDF (ablation) | HiLo · CAPE, SMPL-X noise=0.5 | Chamfer distance: 1.6652 -> 1.3593 | Helped |
Surprising findings
- Adding SDF to PIFu and PaMIR and removing SDF from ICON show that SDF improves detail across baselines, and ICON without SDF degrades sharply under noise.
Limitations
- Relies on accurate-enough SMPL-X estimates; performance degrades with large body noise in baselines, and HiLo's gains are measured on synthetic-noise perturbations.
- Evaluated only on THuman2.0 and CAPE plus 200 in-the-wild images without quantitative metrics on the latter.
Experimental setup
Datasets: THuman2.0 CAPE
Benchmarks: THuman2.0 CAPE
Metrics: Chamfer distance P2S distance Normals error
Baselines: PIFu PIFuHD PaMIR ICON ECON D-IF
Builds on: ICON PIFu PaMIR D-IF SMPL-X
Closest papers by meaning
- VS: Reconstructing Clothed 3D Human from Single Image via Vertex Shift (CVPR 2024 · similarity 0.87)
- MultiGO: Towards Multi-level Geometry Learning for Monocular 3D Textured Human Reconstruction (CVPR 2025 · similarity 0.87)
- SIFU: Side-view Conditioned Implicit Function for Real-world Usable Clothed Human Reconstruction (CVPR 2024 · similarity 0.87)
- DeClotH: Decomposable 3D Cloth and Human Body Reconstruction from a Single Image (CVPR 2025 · similarity 0.87)
- D^3-Human: Dynamic Disentangled Digital Human from Monocular Video (CVPR 2025 · similarity 0.87)
- DLCA-Recon: Dynamic Loose Clothing Avatar Reconstruction from Monocular Videos (AAAI 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.