Python library to extract A. Torralba's GIST descriptor.

This is just a wrapper for Lear's GIST implementation written in C. It supports both Python 2 and Python 3 and was tested under Python 2.7.15 and Python 3.6.6 on Linux.

Following packages must be installed before building and installing lear-gist-python.

$ pip install numpy

FFTW is required to build lear_gist. Please download the source, then build and install like following. (Install guide is here. Please refer for defail.) Make sure --enable-single and --enable-shared options are set to ./configure.

$ ./configure --enable-single --enable-shared
$ make
$ make install

Because:

• lear-gist requires float version FFTW to work with (--enable-single).
• lear-gist-python requires FFTW to be compiled with -fPIC option (--enable-shared).

Download lear_gist

$ ./download-lear.sh

Build and install

$ python setup.py build_ext
$ python setup.py install

If fftw3f is installed in non-standard path (for example, $HOME/local), use -I and -L options:

$ python setup.py build_ext -I $HOME/local/include -L $HOME/local/lib

import gist
import numpy as np

img = ... # numpy array containing an image
descriptor = gist.extract(img)

This sample uses 8 Scene Categories Dataset.

scikit-learn and scikit-image are required.

$ pip install scikit-learn scikit-image

cd sample
sh download-8scene.sh
# Extract GIST features from images in "spatial_envelope_256x256_static_8outdoorcategories" directory and save them into "features" directory
python feature_extraction.py spatial_envelope_256x256_static_8outdoorcategories features
# Train and test a multi-class linear classifier by features in "features" directory
python scene_classification.py features

• img: A numpy array (an instance of numpy.ndarray) which contains an image and whose shape is (height, width, 3).
• nblocks: Use a grid of nblocks * nblocks cells.
• orientations_per_scale: Use len(orientations_per_scale) scales and compute orientations_per_scale[i] orientations for i-th scale.