Research map / 3D vision / Face and shape
4D-DRESS: A 4D Dataset of Real-World Human Clothing With Semantic Annotations
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.
Introduces 4D-DRESS, a real-world 4D dataset of 64 clothed outfits in 520 motion sequences with vertex-level semantic labels, built with a semi-automatic template-free 4D human parsing pipeline, and benchmarks simulation and reconstruction.
Problem
Existing real-world 4D human scan datasets lack accurately segmented garment meshes and semantic labels, and have few loose garments or dynamic motions, while synthetic datasets lack realism.
Approach
A template-free 4D parsing pipeline that fuses multi-view human parser votes, optical-flow label transfer, and SAM mask scores via graph-cut optimization, with a small human-in-the-loop rectification step, to label every vertex of 78k scan frames.
Key results
- Parsing pipeline with PAR+OPT+SAM reaches 98.8% accuracy on outer outfits in BEDLAM versus 71.4% for PAR only and 88-ish for template SMPL+D baseline below 90%.
- Pipeline assigns vertex labels without manual intervention in 96.8% of frames; manual rectification changed 1.5% of vertices within 3.2% of frames.
- Inner and outer outfits show garment-to-body distances up to 7.12 cm and 14.76 cm; ICON/ECON/SiTH do better on inner clothing while end-to-end PIFu/PIFuHD are more stable on outer garments.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add optical flow labels to PAR (ablation) | 4D human parsing pipeline · CLOTH4D Outer | mean accuracy: 0.714 -> 0.942 | Helped |
| Add optical flow labels to PAR (ablation) | 4D human parsing pipeline · BEDLAM Inner | mean accuracy: 0.910 -> 0.963 | Helped |
| Add segmentation-mask scores to PAR+OPT (ablation) | 4D human parsing pipeline · BEDLAM Outer | mean accuracy: 0.942 -> 0.988 | Helped |
| Add segmentation-mask scores to PAR+OPT (ablation) | 4D human parsing pipeline · CLOTH4D Inner | mean accuracy: 0.969 -> 0.995 | Helped |
| Fine-tune material parameters for HOOD (ablation) | HOOD clothing simulation · Lower garments | CD: 2.070 -> 0.924 | Helped |
| Fine-tune material parameters for HOOD (ablation) | HOOD clothing simulation · Dress | CD: 4.292 -> 2.463 | Helped |
| Use SMPL body-model guidance | single-view human reconstruction methods · 4D-DRESS Inner | reconstruction performance | Mixed |
| Use existing video-based human reconstruction methods on outer outfits | Vid2Avatar and SelfRecon · 4D-DRESS Outer | CD: 2.870 -> 3.014 | Hurt |
Surprising findings
- SMPL-body-guided reconstruction methods (ICON, ECON, SiTH) perform better on inner clothing but degrade substantially on outer garments, while end-to-end PIFu and PIFuHD are more stable across both.
- Template-based SMPL+D parsing stays below 90% accuracy on all synthetic outfits, lower than the template-free pipeline.
Limitations
- Pipeline requires substantial computational time.
- Manual rectification and garment mesh extraction need 3D editing expertise and human effort, limiting scalability.
- Dataset covers 32 participants with an average age of 24, limiting subject diversity.
Experimental setup
Datasets: 4D-DRESS CLOTH4D BEDLAM BUFF X-Humans ActorsHQ
Benchmarks: 4D-DRESS clothing simulation 4D-DRESS clothed human reconstruction 4D-DRESS clothing reconstruction 4D-DRESS video-based human reconstruction CLOTH4D BEDLAM
Metrics: accuracy Chamfer Distance normal consistency IoU stretching energy
Baselines: SMPL+D PAR Only LBS PBNS NCS HOOD PIFu PIFuHD PaMIR ICON PHORHUM ECON
Builds on: Graphonomy RAFT Segment Anything HOOD ClothCap
Closest papers by meaning
- FreeCloth: Free-form Generation Enhances Challenging Clothed Human Modeling (CVPR 2025 · similarity 0.88)
- Disco4D: Disentangled 4D Human Generation and Animation from a Single Image (CVPR 2025 · similarity 0.87)
- D^3-Human: Dynamic Disentangled Digital Human from Monocular Video (CVPR 2025 · similarity 0.87)
- DressWild: Feed-Forward Pose-Agnostic Garment Sewing Pattern Generation from In-the-Wild Images (NeurIPS 2026 · similarity 0.87)
- ChatGarment: Garment Estimation, Generation and Editing via Large Language Models (CVPR 2025 · similarity 0.86)
- Garment Recovery with Shape and Deformation Priors (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.