Geometry Aware Convolutional Filters for Omnidirectional Images Representation

Implemtentation of the Geometry Aware Convolutional Filters for Omnidirectional Images Representation ICML 2019 paper

Installation

• to install all the dependencies run:

pip install -r requirements.txt

• add the path to the code to the PYTHONPATH environment variable as shown below:

export PYTHONPATH=<path_to_the_code>:$PYTHONPATH

Usage

This code implements the classification experiment described in the paper. In order to train a model download the dataset, as described in the following section and run

python classification/run_classification.py --exp=<exp_type>

• exp_type - type of the experiment which can be one of the following: [cubic|fisheye|spherical|modspherical]
• by default the experiment with cube-map projection will be executed
• you may need to adjust classification\config in order to run custom experiments

Datasets

Download the datasets as follows:

• Cube-map projection [94.7 MB] -- required for running the code with --exp=cubic

cd data
wget --no-check-certificate -O MNISTcubic.zip https://drive.switch.ch/index.php/s/sVe1wFtqaVRwmqn/download
unzip MNISTcubic.zip
rm MNISTcubic.zip
cd ../

• Fish-eye projection [62.1 MB] -- required for running the code with --exp=fisheye

cd data
wget --no-check-certificate -O fisheye.zip https://drive.switch.ch/index.php/s/WSEy61zestEVyAQ/download
unzip fisheye.zip
rm fisheye.zip
cd ../

• Spherical projection [34.5 MB] -- required for running the code with --exp=spherical

cd data
wget --no-check-certificate -O MNISTomni.zip https://drive.switch.ch/index.php/s/5Kg8DTmhMep3iXi/download
unzip MNISTomni.zip
rm MNISTomni.zip
cd ../

• Modified Spherical projection [131.8 MB] -- required for running the code with --exp=modspherical

cd data
wget --no-check-certificate -O MNISTrandom_projection.zip https://drive.switch.ch/index.php/s/vFsZY38smcu7jA6/download
unzip MNISTrandom_projection.zip
rm MNISTrandom_projection.zip
cd ../

References

If you are using the code please cite the following paper:

@inproceedings{KhasanovaICML19,
	author    = {Reanta Khasanova and Pascal Frossard},
	title     = {Geometry Aware Convolutional Filters for Omnidirectional Images Representation},
	booktitle = {International Conference on Machine Learning},
	year      = {2019}
}