PKU-DyMVHumans Dataset

Overview

PKU-DyMVHumans is a versatile human-centric dataset designed for high-fidelity reconstruction and rendering of dynamic human performances in markerless multi-view capture settings.

It comprises 32 humans across 45 different dynamic scenarios, each featuring highly detailed appearances and complex human motions.

Sources

Project page: https://pku-dymvhumans.github.io
Github: https://github.com/zhengxyun/PKU-DyMVHumans
Paper: https://arxiv.org/abs/2403.16080

Key Features:

High-fidelity performance: We construct a multi-view system to capture humans in motion, containing 56/60 synchronous cameras with 1080P or 4K resolution.
High-detailed appearance: It captures complex cloth deformation, and intricate texture details, like delicate satin ribbon and special headwear.
Complex human motion: It covers a wide range of special costume performances, artistic movements, and sports activities.
Human-object/scene interactions: These include human-object interactions, as well as challenging multi-person interactions and complex scene effects (e.g., lighting, shadows, and smoking).

Benchmark

The objective of our benchmark is to achieve robust geometry reconstruction and novel view synthesis for dynamic humans under markerless and fixed multi-view camera settings, while minimizing the need for manual annotation and reducing time costs.

This includes neural scene decomposition, novel view synthesis, and dynamic human modeling.

Dataset Details

Agreement

Note that by downloading the datasets, you acknowledge that you have read the agreement, understand it, and agree to be bound by them:

The PKU-DyMVHumans dataset is made available only for non-commercial research purposes. Any other use, in particular any use for commercial purposes, is prohibited.
You agree not to further copy, publish or distribute any portion of the dataset.
Peking University reserves the right to terminate your access to the dataset at any time.

Dataset Statistics

Scenes: 45 different dynamic scenarios, engaging in various actions and clothing styles.
Actions: 4 different action types: dance, kungfu, sport, and fashion show.
Individual: 32 professional players, including 16 males, 11 females, and 5 children.
Frames: totalling approximately 8.2 million frames.

Dataset Structure

For each scene, we provide the multi-view images (./case_name/per_view/cam_*/images/), the coarse foreground with RGBA channels (./case_name/per_view/cam_*/images/), as well as the coarse foreground segmentation (./case_name/per_view/cam_*/pha/), which are obtained using BackgroundMattingV2.

To make the benchmarks easier compare with our dataset, we save different data formats (i.e., Surface-SOS, NeuS, NeuS2, Instant-ngp, and 3D-Gaussian) of PKU-DyMVHumans at Part1 and write a document that describes the data process.

.
|--- <case_name>
|   |--- cams                    
|   |--- videos
|   |--- per_view                
|   |--- per_frame              
|   |--- data_ngp       
|   |--- data_NeuS
|   |--- data_NeuS2
|   |--- data_COLMAP
|   |--- <overview_fme_*.png>
|--- ...

BibTeX

@article{zheng2024DyMVHumans,
  title={PKU-DyMVHumans: A Multi-View Video Benchmark for High-Fidelity Dynamic Human Modeling},
  author={Zheng, Xiaoyun and Liao, Liwei and Li, Xufeng and Jiao, Jianbo and Wang, Rongjie and Gao, Feng and Wang, Shiqi and Wang, Ronggang},
  journal={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2024}
}