zhaofang0627 / HPBTT

HPBTT-Pytorch

Human Parsing Based Texture Transfer from Single Image to 3D Human via Cross-View Consistency

Code for the paper Human Parsing Based Texture Transfer from Single Image to 3D Human via Cross-View Consistency (NeurIPS 2020).

Requirements

• Cuda 9.2
• Python 3.6
• Pytorch 1.4
• Opencv 3.4.2

Other requirements:

pip install -r requirements.txt

To install Neural Mesh Renderer and Perceptual loss:

cd ./HPBTT/external
sh install_external.sh

Datasets

• Market-1501
• DeepFashion (In-shop Clothes Retrieval Benchmark)

Demo

• Download the trained models.
• Extract and put the models of Market-1501 and DeepFashion in ./HPBTT/cachedir/snapshots and HMR models in ./HPBTT/external/hmr
• Run the demo:

sh ./HPBTT/run_demo_market.sh <model_name> <epoch_num> <input_image_name>

Training

• Download the Market-1501 dataset and PRW dataset for background.
• Extract and put the datasets in ./dataset
• Generate input images and their SMPL parameters:

python -m HPBTT.experiments.hmr_market

• Train the model:

python -m HPBTT.experiments.train_market --name <model_name>

Evaluation

To evaluate the model:

sh ./HPBTT/run_eval_market.sh <model_name> <epoch_num> <output_image_path>

Acknowledgement

Our code is based on cmr.

Citation

If you use this code for your research, please consider citing:

@inproceedings{zhao2020human,
  title={Human Parsing Based Texture Transfer from Single Image to 3D Human via Cross-View Consistency},
  author={Zhao, Fang and Liao, Shengcai and Zhang, Kaihao and Shao, Ling},
  booktitle={Advances in Neural Information Processing Systems},
  year={2020}
}