APPNet

Attention-based Progressive Partition Network for Human Parsing

This is an efficient implementation of APPNet.

Download

Plesae download LIP dataset.

The trained models can be found at google drive.

Environments

• Python 3.5

• PyTorch 0.4.1

• cffi

• matplotlib

• numpy

• opencv-python

• scipy

• tqdm

• You need to use InPlace-ABN with CUDA implementation, which must be compiled with the following commands:

cd libs
sh build.sh
python build.py

• The model is trained on NVIDIA TITAN 1080 Ti GPU.

Training

• Please set the dataset dir in file 'run.sh'. The contents of each dataset include:

  ─ train_images

  ─ train_segmentations

  ─ val_images

  ─ val_segmentations

  ─ train_id.txt

  ─ val_id.txt

• Please put the pretrained resnet101-imagenet.pth in './dataset/'.

• Run the sh run.sh.

Evaluation

If you want to evaluate the trained models on LIP, you can run the sh run_evaluate.sh.