ZhaoJ9014 / SS-NAN

keras implement of the paper Self-Supervised Neural Aggregation Networks for Human Parsing

SS-NAN

Keras implementation for the CVPR 2017 workshop paper Self-Supervised Neural Aggregation Networks for Human Parsing

This code implements three kinds of models for human parsing dataset LIP
Currently only the re-implementation of the original SS-NAN method available.

Results:

Pixel Accuracy	Mean Accuracy	Mean IoU
85.8%	58.1%	47.90%

Requirments:

keras 2.0.9
tensorflow 1.3.0
python 3.5.4

• Anaconda=5 (not neccessary just for convenience)

Data Preparation

Please download the LIP dataset

Usage:

Evaluation

python LIP.py evaluate --model path_to_model.h5  --dataset  dataset_path/Single_Person --evalnum 0

evalnum=0 uses the whole valset. A positive evalnum indicates the number of images to use for evaluation

Test:

run

demo.py 

to run test on some specific images (the main procedure is to call model.detect())

Train

python LIP.py train --model path_to_model.h5  --dataset  dataset_path/Single_Person  trainmode pretrain

3 kinds of trainmodes available: pretrain, finetune, or fintune_ssloss_withdeep, which correspond to the 3 steps introduced in the paper Self-Supervised Neural Aggregation Networks for Human Parsing

Step1: download pspnet_pretrainweights
set the parameters of model.train()

epochs=40,layers='all'   

run

python LIP.py train --model pspnet  --dataset  dataset_path/Single_Person  trainmode pretrain

Step2 : set the parameters of model.train()

epochs=30,layers='head'  

train the Neural Aggregation Networks

python LIP.py train --model pretain.h5(the best model generated in step1 )  --dataset  dataset_path/Single_Person  trainmode 
finetune

Step3 : set the parameters of model.train()

epochs=30,layers='psp5+'

train with Self-Supervised Loss

python LIP.py train --model finetune.h5(the best model generated in step2 )  --dataset  dataset_path/Single_Person  trainmode finetune_ssloss_withdeep

The final Pretrain_model can be downloaded here

References:

Some codes are borrowed from the MASK RCNN Implementation

@inproceedings{Gong2017Look,
  title={Look into Person: Self-Supervised Structure-Sensitive Learning and a New Benchmark for Human Parsing},
  author={Gong, Ke and Liang, Xiaodan and Zhang, Dongyu and Shen, Xiaohui and Lin, Liang},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  pages={6757-6765},
  year={2017},
}

@inproceedings{Zhao2017Self,
  title={Self-Supervised Neural Aggregation Networks for Human Parsing},
  author={Zhao, Jian and Li, Jianshu and Nie, Xuecheng and Zhao, Fang and Chen, Yunpeng and Wang, Zhecan and Feng, Jiashi and Yan, Shuicheng},
  booktitle={Computer Vision and Pattern Recognition Workshops},
  pages={1595-1603},
  year={2017},
}