Try The Demo

http://101.6.96.191:8501/

Go walking under Van Gogh’s Starry Night!

Installation Requirement

python 3.6
pytorch 1.1.0 （conda install pytorch torchvision torchaudio)

pip install -r requirements.txt

Run The Demo

To try the demo, run streamlit run main.py --server.maxUploadSize 1

One-click run version (including the installation, but might not work properly): bash run.sh

Files

├── main.py
├── README.md
├── make_naive.py
├── auto_localization.py
├── harmonize.py
├── baiduAPI.py
├── requirements.txt
├── run.sh
├── Con_Sin_GAN
│   ├── ConSinGAN
│   ├── evaluate_model.py
│   └── main_train.py
├── Images
│   ├── default_human
│   ├── default_paintings
│   └── tmp
├── TrainedModels
│   ├── nightsky
│   ├── oil_building
│   ├── oil_tree
│   ├── pen_tree
│   ├── scream
│   ├── some_painting
│   ├── street
│   └── watercolor_building
└── MySegmentation
    ├── dataset.py
    ├── gen_mask.py
    ├── loadmodel.py
    ├── log
    ├── models
    ├── padding_data.py
    ├── pics
    ├── print_log
    ├── README.md
    └── segmentation_train.py

Con_Sin_GAN

Con_Sin_GAN is the directory for ConSinGAN implementations obtained from https://github.com/tohinz/ConSinGAN and modified by us for this project.

Images

This directory saves the default painting images and human images, as well as some temporary images.

TrainedModels

Models in this directory corresponds to the default painting images.

MySegmentation

MySegmentation provides a FPN model implemented by ourselves. (But in our final demo, we use BaiduAPI for this project due to its better performance.)

1. Train Model:

Dataset:

Download Supervisely Person Dataset.

Use ./gen_mask.py to extract label from .json file.

Use ./padding_data.py to get images with same size.

Train:

nohup python -u  segmentation_train.py  &> ./log/CVFPN.out&

Log:

Use ./print_log/printlog.ipynb to print the log.

2. Use Model:

Use python loadmodel.py to get the segmentation result.

Azure-Vision/human_in_painting