Iteratively Trained Interactive Segmentation

This is an official TensorFlow implementation for

Sabarinath Mahadevan, Paul Voigtlaender, and Bastian Leibe,
"Iteratively Trained Interactive Segmentation",
British Machine Vision Conference, 2018.

Requirements

Python 3
Tensorflow
Scipy
scikit-image
numpy
pillow

Citation

If you use this code or models, please cite the following:

@inproceedings{mahadevanitis,
  author={Sabarinath Mahadevan and Paul Voigtlaender and Bastian Leibe},
  title={Iteratively Trained Interactive Segmentation},
  booktitle={British Machine Vision Conference (BMVC)},
  year={2018},
}

Pre-trained Models

You can download the pre-trained models from our internal server. All available models are in a single tar.gz file. Currently it contains models that can be used to reproduce the results for iFCN and ITIS in Table 1, and for the ablation study in Figure 5 (see paper for details.

Usage

Download PascalVOC dataset (http://host.robots.ox.ac.uk/pascal/VOC/)
Create a folder data within the source root directory, and copy the Pascal VOC dataset files to it. Alternatively, add a parameter "data_dir: <path to pascal voc root>" in the respective config files.
Download the weights as explained in the previous section, and place them under 'models' directory. Alternatively, you could change the paramter "load" to point it to the required path.
Run the following to evaluate the given models

iFCN:         python main.py configs/pascal_ifcn
iFCN + gauss: python main.py configs/pascal_gauss
ITIS:         python main.py configs/pascal_itis