UltraPose: Synthesizing Dense Pose with 1 Billion Points by Human-body Decoupling 3D Model
Official repository for the ICCV 2021 paper:
UltraPose: Synthesizing Dense Pose with 1 Billion Points by Human-body Decoupling 3D Model [PDF]
Haonan Yan, Jiaqi Chen, Xujie Zhang, Shengkai Zhang, Nianhong Jiao, Xiaodan Liang, Tianxiang Zheng
The dataset is now available at Baidu net disk (code: bpi2) or google drive.
Introduction
In this work, we introduce a new 3D human-body model with a series of decoupled parameters that could freely control the generation of the body. Furthermore, we build a data generation system based on this decoupling 3D model, and construct an ultra dense synthetic benchmark UltraPose, containing around 1.3 billion corresponding points.
Installation
We recommend creating a clean conda environment and install all dependencies. You can do this as follows:
step1
conda create -n ultrapose python=3.7
conda activate ultrapose
step2
conda install pytorch=1.7.1 torchvision cudatoolkit=10.2 -c pytorch
step3
pip install ml-collections opencv-python imgaug visdom pycocotools Cython future h5py
You need to build python3 densepose for evaluation. You can do this as follows:
cd $UltraPoseDir/eval
make
cd $UltraPoseDir/eval/DensePoseData
bash get_eval_data.sh
Training
For single GPU training, please use default configurations by running:
python train.py --dataroot data/ultrapose
Besides, you can also use visdom to monitor the training process.
python -m visdom.server
python train.py --dataroot data/ultrapose --use_visdom
For multi-GPU training with default configurations, you can modify train_transformer.sh
accordingly and run:
sh train_transformer.sh
Evaluation
python evaluation.py
Dataset
The dataset is now available from Baidu net disk (code: bpi2) or google drive.
Extract the data and put them under $UltraPoseDir/data
.
Dataset | Persons | Points | #Avg Density | Mask Resolution | No error |
---|---|---|---|---|---|
Densepose-COCO | 49K | 5.2M | 106 | 256x256 | |
UltraPose | 5K | 13M | 2.6K | 512x512 | β |
Acknowledgements
Parts of the code are taken or adapted from the following repos:
Citation
If you use this code or Ultrapose for your research, please cite our work:
@inproceedings{yan2021ultrapose,
title={UltraPose: Synthesizing Dense Pose With 1 Billion Points by Human-Body Decoupling 3D Model},
author={Yan, Haonan and Chen, Jiaqi and Zhang, Xujie and Zhang, Shengkai and Jiao, Nianhong and Liang, Xiaodan and Zheng, Tianxiang},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
pages={10891--10900},
year={2021}
}