Superpixel-based Graph Convolutional Network for Semantic Segmentation

Citation

@article{jungsuperpixel,
  title={Superpixel-based Graph Convolutional Network for Semantic Segmentation},
  author={Jung, Hoin and Park, Seong Yeon and Yang, Su and Kim, Jin}
}

Introduction

Concept

Thie repository is an implementation for Superpixel-GCN(pdf).
The concept of this framework is regarding Superpixel as a node of an graph.
The proposed method aims classify the label of each superpixel(node) using GCN, and convert graph to image again.

Architecture

Implementation Detail

Dataset

You can download UAVid dataset from here.
Moreover, our framework can be used for various semantic segmentation dataset such as (DeepGlobe Landcover, ADE20K, Cityscapes, etc.)
Note that thie repo includes UAVidToolKit to encode ground truth image to labeled gt image.
Other datasets are necessarily encoded as labeled ground truth.

Superpixel

We are mainly using SFCN as superpixel prerpocessing. However, any other superpixel framework can be adopted.

File tree

|-- data
|   `-- uavid
|       |--uavid_train 
|       |--uavid_val
|       `--uavid_test
|-- preprocessing
|   |-- UAVidToolKit
|   |-- superpixels
|   |   `--superpixel_sfcn
|   |-- uav_preprocessing_sfcn.py 
|   `-- uav_preprocessing_slic.py 
|-- SGNN
|   |-- GraphSage.py  
|   |-- dataloader.py  
|   |-- losses.py  
|   |-- test_eval_uav.ipynb  
|   |-- train.py  
|   |-- utils.py  
|   `-- yml
|       `--train_uav.yaml

Tutorial

Preprocessing

Create superpixel-graph using SFCN.

cd preprocessing
python uav_preprocessing_sfcn.py

Then, *.bin files for train subset and val(test) subset will be created at pickles/.
Also, *.pickle files for val(test) subset is created at pickles/.
*.bin files are used for training and validation(test), while *.pickle file is used for evaluation,
because *.bin file is a list of DGL graphs, where *.pickle contains the superpixel mapping information.

Set configures

Most of options are shown in yml/train_uav.yaml.
You can also use argparser in train.py instead of yaml configure.
yaml works as default options, and you may be able to let argparser override default setting.

Wandb

Be careful our methods contains wandb option.
If you don't want to use wandb, set wandb option in yaml file as False.
Otherwise, please change entity in wandb.init in train.py as your wandb name.

Run Training

cd SGNN
python train.py

Checkpoints are saved every best epochs.
If you want to save memory space, save_best_epoch_pth=False in *.yaml to restore only final.pth.
Note that trin_id will be created automatically following current Unix time.

Test and Evaluation

Run test_eval_uav.ipynb. It also follows configuration of train_uav.yaml.