LinaShanghaitech / CE2P

CE2P

As someone can't reproduce the results with the old version of Pytorch 0.3.1, we reimplement it with Pytorch 0.4.1. 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

Requirements

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'.

Compiling

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.

Dataset and pretrained model

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.

Training and Evaluation

./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:

@article{CE2P2019,
  title={Devil in the Details: Towards Accurate Single and Multiple Human Parsing},
  author={Tao Ruan, Ting Liu, Zilong Huang, Yunchao Wei, Shikui Wei, Yao Zhao, Thomas Huang},
  journal={arXiv:1809.05996},
  year={2018}
}