Instance segmentation

SCHP: Self Correction for Human Parsing

SCHP1 is an out-of-box human parsing representation extractor.

Given an environment of EasyMocap, we can install this code as following:

cd 3rdparty
git clone
cd Self-Correction-Human-Parsing
conda install ninja
python3 -m pip install scikit-image
python3 -m pip install networks

Install detectron2

Install detectron2 according to your cuda version and torch version. For example cu111+torch.19:

python -m pip install detectron2 -f

Download the models

Download the models here

Place the related models in a folder:

├── detectron2_maskrcnn_cihp_finetune.pth
├── exp_schp_multi_cihp_global.pth
└── exp_schp_multi_cihp_local.pth


Given a dataset as EasyMocap format, extract the instance segmentation with this script:

python3 ${data} --ckpt_dir /path/to/models

Results can be found in data/

This code will generate some very big(>100G) outputs. If you don’t have enough space in data/, please add ` –tmp /path/to/tmp`

  1. Li, Peike, et al. “Self-correction for human parsing.” IEEE Transactions on Pattern Analysis and Machine Intelligence (2020).