README.org

#+TITLE: M-CE2P

The codes for generating masks and boxes for MHP v2.0 are now updated in =code_for_maskrcnn= folder. The process for CIHP dataset is similar with that for MHP v2.0.

• Introduction

This respository includes a PyTorch implementation of M-CE2P that won the 1st places of Multi-Person Human Parsing Challenge(track2) and Fine-Grained Multi-Human Human Parsing Challenge(track5) in the 2nd LIP Challenge.

The code is based upon [[https://github.com/liutinglt/CE2P][CE2P]].

• Requirements

  • Pytorch 0.3.1
  • tqdm
  • Pillow
  • python-opencv
  • other packages required by [[https://github.com/liutinglt/CE2P][CE2P]]

• Basic Setup

  Just do what [[https://github.com/liutinglt/CE2P][CE2P]] requires to do :)

• Dataset and Model

  Please download [[http://sysu-hcp.net/lip/overview.php][CIHP]] and [[https://lv-mhp.github.io/dataset][MHP v2.0]]. In the meanwhile, create a folder named =weights= in root folder of this project, download trained model(which are also used in our paper) from [[https://pan.baidu.com/s/1zLWMfDTrPLc9kvis-I21bw][Baidu Drive(Code: yq8f)]] or [[https://drive.google.com/file/d/1JqZ70YFEyIzhjyCgjuOKIItx0Erzj02Y/view?usp=sharing][Google Drive]] and put them into =weights=.

• Training

  To train the model, please modify ./code_for_maskrcnn/MHP_infer_simple_train.py to fit your requirement and follow the step introduced in [[https://github.com/liutinglt/CE2P][CE2P]].

• Evaluation ** Human Mask and Bounding Box Generation We use Mask R-CNN in [[https://github.com/facebookresearch/Detectron][Detectron]] to generate human masks and bounding boxes associated with an input image, and use the default configuration introduced in [[https://github.com/facebookresearch/Detectron/blob/master/GETTING_STARTED.md][Quick Start]].

  For the convenience of users, we also provide the masks and boxes in [[https://pan.baidu.com/s/1_YX2BqcxUihY4XgXcfm66A][Baidu Drive]] and [[https://drive.google.com/open?id=1RF0UN-Csoyc7mkt9txVPKouzKZ2AZol9][Google Drive]], which are generated on validation sets of the two datasets. You can also use script =./code_for_maskrcnn/MHP_infer_simple_test.py= to generate by yourself.

** Run Inference

1. Modify USE_CIHP_DATA to run specific dataset.

2. Modify dataset related paths in =M-CE2P-test.py=.

3. Simply run =python M-CE2P-test.py= to generate the results(stored in =outputs= as default).

** Evaluation

=metrics= folder contains all the metric codes used in our papar.