DTO-Humans
收藏资源简介:
DTO-Humans是一个新的大规模伪真实标注数据集,包含56万张高质量、场景一致的多人物图像,具有密集人群特征,平均每张图像包含4.8个人。该数据集通过将深度条件平移优化(DTO)框架应用于4D-Humans数据集构建而成,利用人体高度先验和单目深度估计器的深度线索,在最大后验概率框架下求解所有受试者的场景一致放置。
DTO-Humans is a novel large-scale pseudo-ground-truth annotated dataset, which consists of 560,000 high-quality, scene-consistent multi-person images featuring dense crowd configurations, with an average of 4.8 individuals per image. This dataset is constructed by applying the Deep Conditional Transformation Optimization (DTO) framework to the 4D-Humans dataset. Leveraging human height priors and depth cues from monocular depth estimators, it solves the scene-consistent placement of all subjects within the maximum a posteriori (MAP) framework.
DTO-Humans 数据集概述
数据集基本信息
- 数据集名称:DTO-Humans
- 论文标题:Towards Metric-Aware Multi-Person Mesh Recovery by Jointly Optimizing Human Crowd in Camera Space
- 论文链接:http://arxiv.org/abs/2511.13282
- 数据集规模:56万张高质量、场景一致的多人物图像
- 平均人物密度:每张图像平均4.8人
数据集特点
- 基于4D-Humans数据集构建
- 采用深度条件平移优化(DTO)方法
- 在相机空间中联合优化人群的平移
- 利用人体高度先验和单目深度估计器的深度线索
- 在原则性最大后验(MAP)框架内求解所有主体的场景一致放置
下载地址
- Google Drive:https://drive.google.com/drive/folders/1ddc43P6iYIctAvmuravIxbxZm3F2uB41?usp=drive_link
- Tsinghua Cloud:https://cloud.tsinghua.edu.cn/d/539173c2952b40f5a422/
数据组织结构
data/ ├── aic/ │ ├── images/ │ ├── aic-release.npz │ └── AIC_CHMR_SMPL_OPT.npz ├── coco2014/ │ ├── images/ │ │ └── train2014/ │ ├── coco-release.npz │ └── COCO_CHMR_SMPL_OPT.npz ├── insta/ │ ├── images/ │ │ └── insta-train/ │ ├── insta1-release.npz │ ├── insta2-release.npz │ └── INSTA_CHMR_SMPL_OPT.npz └── mpii/ ├── images/ ├── mpii-release.npz └── MPII_CHMR_SMPL_OPT.npz
许可证信息
- 许可证类型:Creative Commons Attribution-NonCommercial 4.0 International License
- 使用限制:仅限非商业学术研究用途
引用格式
bibtex @article{wang2025dtohumans, title={Towards Metric-Aware Multi-Person Mesh Recovery by Jointly Optimizing Human Crowd in Camera Space}, author={Kaiwen Wang, Kaili Zheng, Yiming Shi, Chenyi Guo, Ji Wu}, journal={arXiv preprint arXiv:2511.13282}, year={2025} }




