About

A PSPNet(Pyramid Scene Parsing Network) implementation with Tensorflow.

Set up

• Prepare for dataset

  • Download Cityscape from https://www.cityscapes-dataset.com/downloads/

  • Convert labels to trainIds

    reference: https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/preparation/createTrainIdLabelImgs.py

  • Generate filename list

    • Make Cityscape dataset have the following directory

       + Cityscape	
          + leftImg8bit
               + train
               + val
               + test
          + gtFine	
               + train
               + val
               + test
      

    • Config 'CITYSCAPE_DIR' in the cityscape.py
    • python cityscape.py
    • The directory should be as follows after run 'python cityscape.py':

      + Cityscape
         + img_test.txt	
         + img_train.txt	
         + img_val.txt
         + anno_test.txt
         + anno_train.txt	
         + anno_val.txt	
         + leftImg8bit
              + train
              + val
              + test
         + gtFine	
              + train
              + val
              + test
      

• Download the pretrained model

  • download pretrained resnet101 weight from http://download.tensorflow.org/models/resnet_v2_101_2017_04_14.tar.gz
  • download the trained weight from here if you want to inference and evaluate the model.

Exec

• Train

  • for train + val dataset

    python train.py --dataset trainval

  • for train dataset

    python train.py

• Inference [Use your trained model or download checkpoint here]

  • Inference an image in test set randomly

    python predict.py --prediction_on test

  • Inference an image in val set randomly

    python predict.py --prediction_on val

  • Inference an image in train set randomly

    python predict.py --prediction_on train

  • Inference an specified image by file path(or your own image path)

    python predict.py --file_path /Volumes/Samsung_T5/datasets/Cityscape/leftImg8bit_trainvaltest/leftImg8bit/test/berlin/berlin_000270_000019_leftImg8bit.png

• Evaluation [Use your trained model or download checkpoint here]

  • On test set

    python evaluate.py --dataset test

  • On val set

    python evaluate.py --dataset val

Results

Desc	Repo(%)	Repo(%)	Paper(%)
Train set	train	train+val	train+val
mIoU	73.5	74.3	78.4

Prediction images

Pictures in test set.

ZhongGuancun Road in Beijing.

Tensorboard

• cd summary directory and run the following command

  tensorboard --logdir=./

• loss

• learning rate

Reference

• https://github.com/mcordts/cityscapesScripts
• https://github.com/tensorflow/models/tree/master/research/deeplab
• https://github.com/wutianyiRosun/CGNet