This respository includes a PyTorch implementation of CE2P that won the 1st places of single human parsing in the 2nd LIP Challenge. The M-CE2P used for multiple human parsing is provided in https://github.com/RanTaimu/M-CE2P.
The code is based upon https://github.com/speedinghzl/pytorch-segmentation-toolbox, and the data processing is based upon https://github.com/Microsoft/human-pose-estimation.pytorch
python 3.6
PyTorch 0.4.1
To install PyTorch, please refer to https://github.com/pytorch/pytorch#installation.
Or using anaconda: conda env create -f environment.yaml
Or to use Pytorch 1.0, just replace 'libs' with 'modules' in https://github.com/mapillary/inplace_abn, and rename it to 'libs'.
Some parts of InPlace-ABN have a native CUDA implementation, which must be compiled with the following commands:
cd libs
sh build.sh
python build.py
The build.sh
script assumes that the nvcc
compiler is available in the current system search path.
The CUDA kernels are compiled for sm_50
, sm_52
and sm_61
by default.
To change this (e.g. if you are using a Kepler GPU), please edit the CUDA_GENCODE
variable in build.sh
.
Note that the left and right label should be swapped when the label file is flipped.
Plesae download LIP dataset and create symbolic links: ln -s YOUR_LIP_DATASET_DIR dataset/LIP
The contents of LIP Dataset include:
├── train_images
├── train_segmentations
├── val_images
├── val_segmentations
├── test_images
├── train_id.txt
├── val_id.txt
├── test_id.txt
Please download imagenet pretrained resent-101 from baidu drive or Google drive, and put it into dataset folder.
./run.sh
To evaluate the results, please download 'LIP_epoch_149.pth' from baidu drive or Google drive, and put into snapshots directory.
./run_evaluate.sh
The parsing result of the provided 'LIP_epoch_149.pth' is 53.88 without any bells and whistles,
If this code is helpful for your research, please cite the following paper:
@inproceedings{ruan2019devil,
title={Devil in the details: Towards accurate single and multiple human parsing},
author={Ruan, Tao and Liu, Ting and Huang, Zilong and Wei, Yunchao and Wei, Shikui and Zhao, Yao},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={33},
pages={4814--4821},
year={2019}
}