Research map / 3D vision / 3D generation
MVHumanNet: A Large-scale Dataset of Multi-view Daily Dressing Human Captures
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 MVHumanNet, a large-scale multi-view capture dataset of 4,500 identities, 9,000 daily outfits and 645M frames, with annotations, and runs pilot experiments on action recognition, NeRF, and generative models.
Problem
Human-centric 3D vision lacks a large-scale, high-fidelity multi-view human dataset with diverse identities and everyday clothing, unlike object datasets such as Objaverse and MVImgNet.
Approach
Scale multi-view human capture by focusing on everyday clothing and many identities (4,500 people, two outfits each) captured with a 48-camera 12MP system, plus semi-automatic annotation (RVM+SAM masks, OpenPose/triangulation, SMPL/SMPLX fitting) and pilot studies showing data scaling benefits.
Key results
- Skeleton-based action recognition top-1 accuracy rises from ~30% with 1 view to 78.19% with 8 views (FR-Head).
- StyleGAN2 FID improves from 14.05 (3000 subjects) to 7.08 (5500 subjects); GET3D FID from 41.54 to 25.12.
- Fine-tuning IBRNet on MVHumanNet then HuMMan raises PSNR from 28.06 (scratch) to 29.46.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase the number of training views from 1-view to 8-views | Skeleton-based action recognition methods on MVHumanNet · MVHumanNet | Top-1 accuracy: 33.85 -> 78.19 | Helped |
| Increase the number of training views from 1-view to 8-views | Skeleton-based action recognition methods on MVHumanNet · MVHumanNet | Top-5 accuracy: 51.08 -> 92.45 | Helped |
| Increase training outfits from 100 to 5000 | IBRNet · MVHumanNet | PSNR: 26.05 -> 29.00 | Helped |
| Increase training outfits from 100 to 5000 | IBRNet · MVHumanNet | SSIM: 0.9571 -> 0.9706 | Helped |
| Increase training outfits from 100 to 5000 | IBRNet · MVHumanNet | LPIPS: 0.0555 -> 0.0377 | Helped |
| Increase training outfits from 100 to 5000 | GPNeRF · MVHumanNet | PSNR: 23.27 -> 24.69 | Helped |
| Increase training outfits from 100 to 5000 | GPNeRF · MVHumanNet | SSIM: 0.8688 -> 0.8878 | Helped |
| Increase training outfits from 100 to 5000 | GPNeRF · MVHumanNet | LPIPS: 0.2077 -> 0.1961 | Helped |
| Pretrain on MVHumanNet, then finetune on HuMMan | IBRNet · HuMMan test | PSNR: 28.06 -> 29.46 | Helped |
| Pretrain on MVHumanNet, then finetune on HuMMan | GPNeRF · HuMMan test | PSNR: 20.95 -> 21.89 | Helped |
| Increase the number of training subjects from 3000 to 5500 | StyleGAN2 · MVHumanNet | FID: 14.05 -> 7.08 | Helped |
| Increase the number of training subjects from 3000 to 5500 | GET3D · MVHumanNet | FID: 41.54 -> 25.12 | Helped |
Surprising findings
- Generalizable GPNeRF with SMPL prior scored lower PSNR than IBRNet across all data scales.
Limitations
- Pilot experiments use 62% of data and limited hardware, so results are preliminary.
- Annotations such as SMPL fits and masks are automatically generated and not fully verified.
- Data restricted to everyday clothing and indoor capture studio, limiting diversity of interactions.
Experimental setup
Datasets: MVHumanNet HuMMan
Benchmarks: MVHumanNet test split HuMMan
Metrics: top-1 accuracy top-5 accuracy PSNR SSIM LPIPS FID
Baselines: CTR-GCN InfoGCN FR-Head IBRNet GPNeRF
Builds on: Stable Diffusion StyleGAN2 GET3D IBRNet GP-NeRF SAM RVM OpenPose
Closest papers by meaning
- PKU-DyMVHumans: A Multi-View Video Benchmark for High-Fidelity Dynamic Human Modeling (CVPR 2024 · similarity 0.88)
- HumanVid: Demystifying Training Data for Camera-controllable Human Image Animation (NeurIPS 2024 · similarity 0.87)
- OpenHumanVid: A Large-Scale High-Quality Dataset for Enhancing Human-Centric Video Generation (CVPR 2025 · similarity 0.87)
- BEDLAM2.0: Synthetic humans and cameras in motion (NeurIPS 2025 · similarity 0.87)
- ChatHuman: Chatting about 3D Humans with Tools (CVPR 2025 · similarity 0.86)
- HOI-M^3: Capture Multiple Humans and Objects Interaction within Contextual Environment (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.