Abstract:A large number of diverse data sets are necessary for networks to predict human body parameters and reconstruct 3D body models from images. Due to the high cost of motion capture and body scanning, high precision pose and body shape parameters are difficult to obtain. Meanwhile, existing datasets cannot meet the requirements in terms of diversity, size, and data accuracy for practical applications. Inspired by the construction schemes of various datasets, we design and construct a large multi-view 3D human body reconstruction dataset (3DMVHumanBP) with more types of supervised data. By recording the different poses of 25 women and 25 men in a green screen laboratory from six perspectives, we constructed a complete large multi-view 3D body posture dataset containing 340, 000 images. It is worth noting that, we innovatively propose a body dimension prior to the constrained human parametric model construction strategy to provide high-precision ground truth parameters of the human body SMPL models. In addition, we also designed a dense UV data generation method based on human body boundary and mask mapping to provide high-quality dense UV data, which more closely fits the features of the human images. It makes up for the defect that few existing data sets can only provide sparse UV data. In the experiment, the effectiveness and advantages of the data set constructed by us in network training are verified. Compared with the training of existing datasets, the mainstream network models trained on our datasets can significantly improve their prediction accuracy and robustness, thanks to the monitoring data of multiple kinds of high-precision human model parameters provided by 3DMVHumanBP. We hope that the human body dataset construction scheme we designed can provide ideas for building large-scale high precision human body datasets in the future.

Multi-view 3D Human Physique Dataset Construction for Robust Digital Human Modeling of Natural Scenes.

High-precision Human Body Acquisition Via Multi-View Binocular Stereopsis

A Robust Multi‐View System for High‐Fidelity Human Body Shape Reconstruction

Human3.6M: Large Scale Datasets and Predictive Methods for 3D Human Sensing in Natural Environments

MVHumanNet: A Large-scale Dataset of Multi-view Daily Dressing Human Captures

HuMMan: Multi-Modal 4D Human Dataset for Versatile Sensing and Modeling

3D Body Shapes Estimation from Dressed-Human Silhouettes.

SynBody: Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling

Three-Dimensional Human Body Reconstruction Using Dual-View Normal Maps

3D Human Reconstruction from A Single Depth Image

Unsupervised Universal Hierarchical Multi-Person 3D Pose Estimation for Natural Scenes

Human-M3: A Multi-view Multi-modal Dataset for 3D Human Pose Estimation in Outdoor Scenes

UltraPose: Synthesizing Dense Pose with 1 Billion Points by Human-body Decoupling 3D Model

MVP-Human Dataset for 3D Human Avatar Reconstruction from Unconstrained Frames

H3WB: Human3.6M 3D WholeBody Dataset and Benchmark

MUC: Mixture of Uncalibrated Cameras for Robust 3D Human Body Reconstruction

PKU-DyMVHumans: A Multi-View Video Benchmark for High-Fidelity Dynamic Human Modeling

Multi-view Shape Generation for a 3D Human-like Body

Learning 3-D Human Pose Estimation from Catadioptric Videos

FreeMan: Towards Benchmarking 3D Human Pose Estimation under Real-World Conditions