eDN-saliency

This repository contains reference code for computing Ensembles of Deep Networks (eDN) saliency maps based on the CVPR'2014 paper "Large-Scale Optimization of Hierarchical Features for Saliency Prediction in Natural Images".

Usage

./eDNsaliency [--opts] <image> <output_saliency_map>

Options:
  -h, --help         show this help message and exit
  --descs DESCPATH   path to eDN model(s) (default: ./slmBestDescrCombi.pkl)
  --svm SVMPATH      path to SVM file (default: ./svm-slm-cntr)
  --white WHITEPATH  path to whitening parameters (default: ./whiten-slm-cntr)
  --fixMap FIXMAP    fixation map / empirical saliency map, if available
  --histeq           histogram equalization with given empirical saliency map
                     (default: False); requires empirical saliency map
  --auc              computes AUC for given fixation map; requires fixation map
  --no-blur          disable the default smoothing of the final map

Input format:

• fixation map: black image with fixated pixels (one per fixation) set to 255 (see ./img_fixPts.jpg)
• empirical saliency map: superposition of Gaussians centered at fixations (see ./img_fixMap.jpg)

These images should have the same size as the input image.

Examples

./eDNsaliency img.jpg salMap.jpg 
    Computes raw (non-histogram-equalized) saliency map (in salMap.jpg) for given image 

./eDNsaliency --histeq --fixMap img_fixMap.jpg img.jpg salMap-histeq.jpg
    Computes histogram-equalized saliency map with given empirical saliency map (img_fixMap.jpg)

./eDNsaliency --auc --fixMap img_fixPts.jpg img.jpg salMap.jpg
    Computes Area Under the Curve (AUC) for fixation map (img_fixPts.jpg)
    
./eDNsaliency --svm ./svm-slm --white ./whiten-slm  img.jpg  salMap-noCntr.jpg
    Computes non-centered saliency maps

Requirements

sthor
liblinear

Installation

(Tested under Ubuntu 14.04)

1. Install dependencies

sudo apt-get install python-matplotlib python-setuptools curl python-dev libxml2-dev libxslt-dev

2. Install liblinear

Download toolbox from http://www.csie.ntu.edu.tw/~cjlin/liblinear/

# extract the zip
make
cd python
make

3. Install sthor dependencies

curl -O http://python-distribute.org/distribute_setup.py
sudo python distribute_setup.py
sudo easy_install pip
sudo easy_install -U scikit-image
sudo easy_install -U cython
sudo easy_install -U numexpr
sudo easy_install -U scipy

For speedup, numpy and numexpr should be built against e.g. Intel MKL libraries.

4. Install sthor

git clone https://github.com/nsf-ri-ubicv/sthor.git
cd sthor/sthor/operation
sudo make
cd ../..
python setup.py install

add the sthor directory and the liblinear/python directory to your PYTHONPATH

5. Test sthor installation

python
import sthor  # should import without errors

Precomputed Saliency Maps

We provide precomputed saliency maps for three standard benchmarks:

• MIT data set: MIT1003_eDN.zip
• Toronto data set: Toronto_eDN.zip
• NUSEF data set: NUSEF_eDN.zip

Citing this Code

If you use this code in your own work, please cite the following paper:

Eleonora Vig, Michael Dorr, David Cox, "Large-Scale Optimization of Hierarchical Features for Saliency Prediction in Natural Images", IEEE Computer Vision and Pattern Recognition (CVPR), 2014.

Link to the paper: http://coxlab.org/pdfs/cvpr2014_vig_saliency.pdf

For questions and feedback please contact me at eleonora.vig@dlr.de