iHOG: Inverting Histograms of Oriented Gradients
This software package contains tools to invert and visualize HOG features. It implements the Paired Dictionary Learning algorithm described in our paper "HOGgles: Visualizing Object Detection Features" [1].
Installation
Before you can use this tool, you must compile iHOG. Execute the 'compile' script in MATLAB to compile the HOG feature extraction code and sparse coding SPAMS toolbox:
$ cd /path/to/ihog
$ matlab
>> compile
If you run into trouble compiling the SPAMS code, you might try opening
the file /path/to/ihog/spams/compile.m and adjusting the settings for
your computer.
Remember to also adjust your path so MATLAB can find iHOG:
>> addpath(genpath('/path/to/ihog'))
If you want to use iHOG in your own project, you can simply drop the iHOG directory into the root of your project.
In order to use iHOG, you must have a learned paired dictionary. By default, iHOG will attempt to download a pretrained one from MIT for you on the first execution. If you wish to download it manually, simply do:
$ wget http://people.csail.mit.edu/vondrick/pd.mat
Inverting HOG
To invert a HOG point, use the 'invertHOG()' function:
>> feat = features(im, 8);
>> ihog = invertHOG(feat);
>> imagesc(ihog); axis image;
Computing the inverse should take no longer than a second for a typical sized image on a modern computer. (It may slower the first time you invoke it as it caches the paired dictionary from disk.)
Learning
We provide a prelearned dictionary in 'pd.mat', but you can learn your own if you wish. Simply call the 'learnpairdict()' function and pass it a directory of images:
>> pd = learnpairdict('/path/to/images/', 1000000, 1000, 5, 5);
The above learns a 5x5 HOG patch paired dictionary with 1000 elements and a training set size of one million window patches. Depending on the size of the problem, it may take minutes or hours to complete.
Bundled Libraries
The iHOG package contains source code from the SPAMS sparse coding toolbox (http://spams-devel.gforge.inria.fr/). We have modified their code to better support 64 bit machines.
In addition, we have included a select few files from the discriminatively trained deformable parts model (http://people.cs.uchicago.edu/~rbg/latent/). We use their HOG computation code and glyph visualization code.
Questions and Comments
If you have any feedback, please write to Carl Vondrick at vondrick@mit.edu.
References
If you use our software, please cite our conference paper:
[1] Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba. "HOGgles: Visualizing Object Detection Features" International Conference on Computer Vision (ICCV), Sydney, Australia, December 2013.