Pysaliency is a python package for saliency modelling. It aims at providing a unified interface to both the traditional saliency maps used in saliency modeling as well as probabilistic saliency models.
Pysaliency can evaluate most commonly used saliency metrics, including AUC, sAUC, NSS, CC image-based KL divergence, fixation based KL divergence and SIM for saliency map models and log likelihoods and information gain for probabilistic models.
You can install pysaliency from pypi via
pip install pysaliency
import pysaliency
dataset_location = 'datasets'
model_location = 'models'
mit_stimuli, mit_fixations = pysaliency.external_datasets.get_mit1003(location=dataset_location)
aim = pysaliency.AIM(location=model_location)
saliency_map = aim.saliency_map(mit_stimuli.stimuli[0])
plt.imshow(saliency_map)
auc = aim.AUC(mit_stimuli, mit_fixations)
If you already have saliency maps for some dataset, you can import them into pysaliency easily:
my_model = pysaliency.SaliencyMapModelFromDirectory(mit_stimuli, '/path/to/my/saliency_maps')
auc = my_model.AUC(mit_stimuli, mit_fixations)
Check out the Tutorial for a more detailed introduction!
Pysaliency provides several important datasets:
- MIT1003
- MIT300
- CAT2000
- Toronto
- Koehler
- iSUN
- SALICON (both the 2015 and the 2017 edition and each with both the original mouse traces and the inferred fixations)
- FIGRIM
- OSIE
- NUSEF (the part with public images)
and some influential models:
- AIM
- SUN
- ContextAwareSaliency
- BMS
- GBVS
- GBVSIttiKoch
- Judd
- IttiKoch
- RARE2012
- CovSal
These models are using the original code which is often matlab. Therefore, a matlab licence is required to make use of these models, although quite some of them work with octave, too (see below).
pysaliency will fall back to octave if no matlab is installed.
Some models might work with octave, e.g. AIM and GBVSIttiKoch. In Debian/Ubuntu you need to install
octave
, octave-image
, octave-statistics
, liboctave-dev
.
These models and dataset seem to work with octave:
- models
- AIM
- GBVSIttiKoch
- datasets
- Toronto
- MIT1003
- MIT300
- SALICON
The Judd Model needs some libraries to work. In ubuntu/debian you need to install these packages:
libopencv-core-dev, libopencv-flann-dev, libopencv-imgproc-dev, libopencv-photo-dev, libopencv-video-dev, libopencv-features2d-dev, libopencv-objdetect-dev, libopencv-calib3d-dev, libopencv-ml-dev, opencv2/contrib/contrib.hpp