This repository includes code which converts datasets i.e. Playing for data, into tfrecords so that it will be acceptable by DeepLab project.

Copy the content of this repository in to folder: tensorflow/models/research/deeplab/datasets, except densecrf_processor.py and vis.py.

Download data i.e. from here. Extract them in tensorflow/models/research/deeplab/datasets such that you have following directory structure.

+ datasets
  + gta
    + images
    + labels

# From the tensorflow/models/research/deeplab/datasets directory.
sh convert_gta_pfd.sh.sh

In order to run DenseCRF, run python densecrf_processor.py -e /home/mumu01/exps/exp42/vis/ -n 42. Where -e contains the path to the directory of the results of vis.py and -n contains a different number so that ppm files dont conflict incase of multithreading of this script.

vis.py is a bit modified version of original deeplab/vis.py.

+ creates new folders.
+ breaks the loop in order to check for more checkpoint files.

Place vis.py in deeplab/ directory.