Research map / 3D vision / Face and shape
D^3-Human: Dynamic Disentangled Digital Human from Monocular Video
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 decoupled clothing and body meshes from monocular video by combining an SDF-based human manifold field (hmSDF) for visible regions with SMPL completion for occluded body parts, plus MLP non-rigid deformation.
Problem
Reconstructing separable, detailed clothing and underlying body geometry from a single monocular video, where body parts occluded by clothing are invisible and must be plausibly completed.
Approach
Defines a human manifold signed distance field (hmSDF) on a watertight clothed-human SDF that segments visible clothing from visible body using only 2D human parsing, then fills invisible body regions from SMPL and jointly refines both layers with separate deformation MLPs, LBS, and perceptual normal loss.
Key results
- On four synthetic SelfRecon sequences, D3-Human reports full-body Chamfer Distance of 0.959, 0.636, 0.270, 0.279 (e-3), the best among compared methods on all sequences.
- Reconstruction takes about 20 minutes for a decoupled template, versus over a day for prior methods such as SelfRecon, REC-MV and GoMAvatar.
- Ablation shows hmSDF yields a cleaner garment shape than an implicit UDF, which produces holes and fails to open cuffs under single-view supervision.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use hmSDF instead of implicit UDF for clothing reconstruction (ablation) | D3-Human · male-3-casual, PeopleSnapshot | clothing reconstruction quality | Helped |
| Use perceptual normal loss instead of cosine or MSE loss (ablation) | D3-Human · normal-loss ablation | reconstruction smoothness and detail | Helped |
| Replace perceptual normal loss with cosine loss (ablation) | D3-Human · normal-loss ablation | reconstruction smoothness | Hurt |
| Replace perceptual normal loss with MSE loss (ablation) | D3-Human · normal-loss ablation | reconstruction smoothness | Hurt |
Limitations
- Relies on the accuracy of per-frame human parsing masks, whose errors cause boundary and fragment artifacts requiring region aggregation.
- Occlusion between clothing and body (e.g., hands) can produce incorrect masks and holes in the abdomen region.
- Evaluation on real data is only qualitative because no public real dataset with decoupled ground truth exists.
Experimental setup
Datasets: PeopleSnapshot SelfRecon synthetic dataset
Benchmarks: SelfRecon synthetic sequences (Female1, Female3, Male1, Male2) PeopleSnapshot
Metrics: Chamfer Distance
Baselines: REC-MV BCNet DELTA SelfRecon GoMAvatar
Builds on: DMTet G-Shell SMPL SelfRecon Sapiens
Closest papers by meaning
- DLCA-Recon: Dynamic Loose Clothing Avatar Reconstruction from Monocular Videos (AAAI 2024 · similarity 0.90)
- VS: Reconstructing Clothed 3D Human from Single Image via Vertex Shift (CVPR 2024 · similarity 0.89)
- Link to the Past: Temporal Propagation for Fast 3D Human Reconstruction from Monocular Video (CVPR 2025 · similarity 0.88)
- DeClotH: Decomposable 3D Cloth and Human Body Reconstruction from a Single Image (CVPR 2025 · similarity 0.88)
- Diffusion-FOF: Single-View Clothed Human Reconstruction via Diffusion-Based Fourier Occupancy Field (CVPR 2024 · similarity 0.88)
- MultiGO: Towards Multi-level Geometry Learning for Monocular 3D Textured Human Reconstruction (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.