SIN: Superpixel Interpolation Network

This is is a PyTorch implementation of the superpixel segmentation network introduced in our PRICAI-2021 paper:

SIN: Superpixel Interpolation Network

Prerequisites

The training code was mainly developed and tested with python 3.6, PyTorch 1.4, CUDA 10, and Ubuntu 18.04.

Demo

The demo script run_demo.py provides the superpixels with grid size 16 x 16 using our pre-trained model (in /pretrained_ckpt). Please feel free to provide your own images by copying them into /demo/inputs, and run

python run_demo.py --data_dir=./demo/inputs --data_suffix=jpg --output=./demo 

The results will be generate in a new folder under /demo called spixel_viz.

Data preparation

To generate training and test dataset, please first download the data from the original BSDS500 dataset, and extract it to . Then, run

cd data_preprocessing
python pre_process_bsd500.py --dataset=
   
     --dump_root=
    
     
python pre_process_bsd500_ori_sz.py --dataset=
     
       --dump_root=
      
       
cd ..

      
     
    
   

The code will generate three folders under the , named as /train, /val, and /test, and three .txt files record the absolute path of the images, named as train.txt, val.txt, and test.txt.

Training

Once the data is prepared, we should be able to train the model by running the following command

python main.py --data=
   
     --savepath=
    

    
   

if we wish to continue a train process or fine-tune from a pre-trained model, we can run

python main.py --data=
   
     --savepath=
    
      --pretrained=
      

     
    
   

The code will start from the recorded status, which includes the optimizer status and epoch number.

The training log can be viewed from the tensorboard session by running

tensorboard --logdir=
   
     --port=8888

   

Testing

We provide test code to generate: 1) superpixel visualization and 2) the.csv files for evaluation.

To test on BSDS500, run

python run_infer_bsds.py --data_dir=
   
     --output=
    
      --pretrained=
     

     
    
   

To test on NYUv2, please follow the intruction on the superpixel benchmark to generate the test dataset, and then run

python run_infer_nyu.py --data_dir=
   
     --output=
    
      --pretrained=
     

     
    
   

To test on other datasets, please first collect all the images into one folder , and then convert them into the same format (e.g. .png or .jpg) if necessary, and run

python run_demo.py --data_dir=
   
     --data_suffix=
    
      --output=
     
       --pretrained=
      

      
     
    
   

Superpixels with grid size 16 x 16 will be generated by default. To generate the superpixel with a different grid size, we simply need to resize the images into the approporate resolution before passing them through the code. Please refer to run_infer_nyu.py for the details.

Evaluation

We use the code from superpixel benchmark for superpixel evaluation. A detailed instruction is available in the repository, please

(1) download the code and build it accordingly;

(2) edit the variables $SUPERPIXELS, IMG_PATH and GT_PATH in /eval_spixel/my_eval.sh,

(3) run

cp /eval_spixel/my_eval.sh 
   
    /examples/bash/
cd  
    
     /examples/
bash my_eval.sh

    
   

several files should be generated in the map_csv folders in the corresponding test outputs;

(4) run

cd eval_spixel
python copy_resCSV.py --src=
   
     --dst=
    

    
   

(5) open /eval_spixel/plot_benchmark_curve.m , set the our1l_res_path as and modify the num_list according to the test setting. The default setting is for our BSDS500 test set.;

(6) run the plot_benchmark_curve.m, the ASA Score, CO Score, and BR-BP curve of our method should be shown on the screen. If you wish to compare our method with the others, you can first run the method and organize the data as we state above, and uncomment the code in the plot_benchmark_curve.m to generate a similar figure shown in our papers.

Acknowledgement

The code is implemented based on superpixel_fcn. We would like to express our sincere thanks to the contributors.

Cite

If you use SIN in your work please cite our paper:

@article{yuan2021sin,
title={SIN: Superpixel Interpolation Network},
author={Qing Yuan, Songfeng Lu, Yan Huang, Wuxin Sha},
booktitle={PRICAI},
year={2021}
}