Implemented Architecture of Pyramid Scene Parsing Network in Keras.

1. Install dependencies:
   • Tensorflow (-gpu)
   • Keras
   • numpy
   • scipy
   • pycaffe(PSPNet)(optional for converting the weights)

   pip install -r requirements.txt --upgrade

2. Converted trained weights are needed to run the network. Weights(in .h5 .json format) have to be downloaded and placed into directory weights/keras

Already converted weights can be downloaded here:

• pspnet50_ade20k.h5 pspnet50_ade20k.json
• pspnet101_cityscapes.h5 pspnet101_cityscapes.json
• pspnet101_voc2012.h5 pspnet101_voc2012.json

Running this needs the compiled original PSPNet caffe code and pycaffe.

python weight_converter.py <path to .prototxt> <path to .caffemodel>

python pspnet.py -m <model> -i <input_image>  -o <output_path>
python pspnet.py -m pspnet101_cityscapes -i example_images/cityscapes.png -o example_results/cityscapes.jpg
python pspnet.py -m pspnet101_voc2012 -i example_images/pascal_voc.jpg -o example_results/pascal_voc.jpg

List of arguments:

 -m --model        - which model to use: 'pspnet50_ade20k', 'pspnet101_cityscapes', 'pspnet101_voc2012'
    --id           - (int) GPU Device id. Default 0
 -s --sliding      - Use sliding window
 -f --flip         - Additional prediction of flipped image
 -ms --multi_scale - Predict on multiscale images

• The interpolation layer is implemented as custom layer "Interp"
• Forward step takes about ~1 sec on single image
• Memory usage can be optimized with:

  config = tf.ConfigProto()
  config.gpu_options.per_process_gpu_memory_fraction = 0.3 
  sess = tf.Session(config=config)

• ndimage.zoom can take a long time