Research map / 3D vision / NeRF and radiance fields
PKU-DyMVHumans: A Multi-View Video Benchmark for High-Fidelity Dynamic Human Modeling
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 PKU-DyMVHumans, a multi-view video dataset of 32 performers across 45 dynamic scenarios (8.2M frames), plus an off-the-shelf NeRF/SDF benchmark pipeline for novel view synthesis, dynamic human modeling, and neural scene decomposition.
Problem
Existing multi-view human datasets lack high-fidelity appearance (loose/oversized clothing), complex poses, and multi-person or human-object interaction data needed for high-quality dynamic human reconstruction and rendering.
Approach
A large-scale dense multi-view capture dataset (56/60 synchronized cameras, 1080P/4K) with COLMAP-calibrated camera conversion tools and a unified benchmark pipeline for NeRF/SDF-based methods (Instant-NGP, NeuS, NeuS2, Tensor4D, Surface-SOS).
Key results
- Instant-NGP achieves the highest average novel view synthesis quality on the benchmark (PSNR 32.02, SSIM 0.987, LPIPS 0.019), versus NeuS2 (30.75) and NeuS (23.93).
- Surface-SOS refinement improves foreground segmentation over BGMv2 (mIoU 0.908 vs 0.881; accuracy 0.962 vs 0.938 average).
- In dynamic reconstruction with dense views on long sequences, Instant-NGP averages PSNR 29.82 while NeuS2 averages 28.81; with short sequences NeuS2 reaches 31.20 PSNR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Instant-NGP instead of NeuS | PKU-DyMVHumans novel view synthesis · PKU-DyMVHumans | PSNR: 23.93 -> 32.02 | Helped |
| Use Instant-NGP instead of NeuS | PKU-DyMVHumans novel view synthesis · PKU-DyMVHumans | SSIM: 0.977 -> 0.987 | Helped |
| Use Instant-NGP instead of NeuS | PKU-DyMVHumans novel view synthesis · PKU-DyMVHumans | LPIPS: 0.027 -> 0.019 | Helped |
| Use hash encoding instead of NeuS's representation | PKU-DyMVHumans novel view synthesis · PKU-DyMVHumans | PSNR: 23.93 -> 32.02 | Helped |
| Use Instant-NGP instead of NeuS2 | PKU-DyMVHumans long sequence, dense view · PKU-DyMVHumans | PSNR: 28.81 -> 29.82 | Helped |
| Use Instant-NGP instead of NeuS2 | PKU-DyMVHumans long sequence, dense view · PKU-DyMVHumans | SSIM: 0.982 -> 0.984 | Helped |
| Use Instant-NGP instead of NeuS2 | PKU-DyMVHumans long sequence, dense view · PKU-DyMVHumans | LPIPS: 0.024 -> 0.023 | Helped |
| Apply Surface-SOS refinement instead of BGMv2 alone | PKU-DyMVHumans foreground segmentation · PKU-DyMVHumans | MSE: 0.422 -> 0.224 | Helped |
| Apply Surface-SOS refinement instead of BGMv2 alone | PKU-DyMVHumans foreground segmentation · PKU-DyMVHumans | mIoU: 0.881 -> 0.908 | Helped |
| Apply Surface-SOS refinement instead of BGMv2 alone | PKU-DyMVHumans foreground segmentation · PKU-DyMVHumans | Acc.: 0.938 -> 0.962 | Helped |
Surprising findings
- NeuS2 achieves better average PSNR than Instant-NGP on some scenes in the dynamic setting, but Instant-NGP is best overall for static novel view synthesis, and NeuS2 gives blurrier results despite higher average performance than NeuS.
Limitations
- Benchmarks focus on NeRF/SDF methods; multi-person modeling and complex scene effects remain challenging and are not fully addressed.
- Free-viewpoint rendering from monocular self-rotating video yields artifacts due to occlusion and view absence.
Experimental setup
Datasets: PKU-DyMVHumans
Benchmarks: PKU-DyMVHumans
Metrics: PSNR SSIM LPIPS mIoU MSE accuracy
Baselines: NeuS Instant-NGP NeuS2 Tensor4D BGMv2 Surface-SOS
Builds on: NeRF NeuS Instant-NGP NeuS2 Tensor4D
Closest papers by meaning
- MVHumanNet: A Large-scale Dataset of Multi-view Daily Dressing Human Captures (CVPR 2024 · similarity 0.88)
- ODHSR: Online Dense 3D Reconstruction of Humans and Scenes from Monocular Videos (CVPR 2025 · similarity 0.88)
- HumanVid: Demystifying Training Data for Camera-controllable Human Image Animation (NeurIPS 2024 · similarity 0.88)
- Reconstructing People, Places, and Cameras (CVPR 2025 · similarity 0.87)
- HyperHuman: Hyper-Realistic Human Generation with Latent Structural Diffusion (ICLR 2024 · similarity 0.87)
- SyncHuman: Synchronizing 2D and 3D Generative Models for Single-view Human Reconstruction (NeurIPS 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.