pytorch-polygon-rnn

Pytorch implementation of Polygon-RNN. Notice that I use another way to handle the first vertex instead of training another model as in the paper. I will not maintain the project, please refer to Polygon-RNN++ for better experience.

Difference with the original paper

1. I use two virtual starting vertices to handle the first vertex as in the image captioning.

2. I add a LSTM layer after the ConvLSTM layers since I need the output to be D*D+1 dimension to handle the end symbol.

How to train and test

1. Download data from CityScapes, organize the image files and annotation json files as follows:

img
├── train
│   ├── cityname1
│   │   ├── pic.png
│   │   ├── ...
│   ├── cityname2
│   │   ├── pic.png
│   │   ├── ...
├── val
│   ├── cityname
│   │   ├── pic.png
│   │   ├── ...
├── test
│   ├── cityname
│   │   ├── pic.png
│   │   ├── ...

label
├── train
│   ├── cityname1
│   │   ├── annotation.json
│   │   ├── ...
│   ├── cityname2
│   │   ├── annotation.json
│   │   ├── ...
├── val
│   ├── cityname
│   │   ├── annotation.json
│   │   ├── ...
├── test
│   ├── cityname
│   │   ├── annotation.json
│   │   ├── ...

The png files and the json files should have corresponding same name.

Execute the following command to make directories for new data and save models:

mkdir -p new_img/(train/val/test)
mkdir -p new_label/(train/val/test)
mkdir save

2. Run the following command to generate data for train/validation/test.

python generate_data.py --data train/val/test

3. Run the following command to train.

python train.py --gpu_id 0 1 2 --batch_size 8  --lr 0.0001

4. Run the following command to validate.

python validate.py --gpu_id 0 1 2 --batch_size 8

5. Run the following command to test.

python test.py --gpu_id 0 --model ./save/model.pth

Now you can easily change configurations on default_config.yaml.