This software package contains code used for conducting experiments in following work:
A. Bendale, T. Boult โTowards Open Set Deep Networksโ IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016 pdf
Authors: Abhijit Bendale ([email protected]) Terrance Boult ([email protected]) Vision and Security Technology Lab University of Colorado at Colorado Springs
Citing the paper
@InProceedings{bendale-boult-cvpr2016,
title={Towards Open Set Deep Networks},
author={Bendale, Abhijit and Boult, Terrance},
booktitle={Computer Vision and Pattern Recognition (CVPR), 2016 IEEE Conference on},
year={2016},
organization={IEEE}
}
The code is provided "as is", without any guarantees. Please refer COPYRIGHT.txt and libMR/COPYRIGHT_Libmr.txt for more details about license and usage restrictions.
The package is divided into two parts: LibMR and OpenMax.
-
LibMR contains code for Extreme Value Theory based Weibull fitting (and other code used in the work for Meta-Recognition). Weibull fitting performed by LibMR is slightly different than that done by other popular packages such as MATLAB's wblfit() function. We have added a short tutorial to illustrate these differences.
-
OpenMax contains code used in the CVPR 2016 paper on "Towards Open Set Deep Networks". The software packages for OpenMax code calls LibMR functions for performing Weibull fitting. In this code, we are providing code to compute OpenMax probabilities from pre-computed Mean Activation Vectors and Softmax probabilities. Mean Activation Vectors and Softmax probabilities are obtained by passing images through Caffe's architecture.
Compile LibMR and python interface to LibMR using following commands. For pythong interfaces to work, you would require Cython to be pre-installed on your machine
cd libMR/
chmod +x compile.sh
./compile.sh
are saved in a mat file. We save fc7, fc8, prob and scores from Caffe framework. Example of saved files can be found in data/train_features/
E.g. loading the file data/train_features/n01440764/n01440764_9981.JPEG.mat in ipython will lead to
>>> from scipy.io import loadmat
>>> import scipy as sp
>>> features = loadmat('data/train_features/n01440764/n01440764_9981.JPEG.mat')
>>> print features.keys()
['fc7', 'fc8', '__header__', '__globals__', 'scores', 'IMG_NAME', '__version__', 'prob']
>>> print features['fc8'].shape
(10, 1000)
>>> print sp.mean(features['prob'], axis=0)[0]
0.999071
>>> print features['scores'][0][0]
0.9990708
where fc7, fc8, prob, scores are outputs of respective layers provided by Caffe Library. Further, fc8 layer contains 10 channels for 1000 classes of ImageNet. These channels are referred to as crops in Caffe library classifier.py, check the function predict() ). The average of prob layer (i.e. SoftMax layer) is the final probability value reported by AlexNet as architected in Caffe Library. In the paper (and code), each of these 10 crops is referred to as channel. Hence, average of features['prob'], is essentially features['scores'].
You will have to install Caffe library on your machine with python interface activated. If you do have it configured, then features for entire ImageNet data can be extracted using the script, python preprocessing/imageNet_Features.py
Mean Activation Vector is defined as the Mean Vector in the pen-ultimate layer. In our paper, we considered fc8 as the pen-ultimate layer. Hence, MAV is the mean of features in fc8. The Mean is computed for each channel. The script to compute mean can be found in
python preprocessing/MAV_Compute.py
Since this process is very time consuming, we are proving pre-computed mean activation vectors on ILSVRC 2012 dataset with this code repository. You can download the mean activation vectors using following command
wget http://vast.uccs.edu/OSDN/data.tar
tar -xvf data.tar
The structure of Mean Files is very simple. For e.g.
>>> from scipy.io import loadmat
>>> mav = loadmat('data/mean_files/n01440764.mat')
>>> print mav['n01440764'].shape
(10, 1000)
The above is the MAV for object category n01440764. MAV is computed in similar manner for each of the 1000 ImageNet Categories. Note, for computing MAV, only those images are considered that classified correctly by the network.
Once the mean activation vector for each category is computed, the next step is to compute category specific distance distribution.
>>> from scipy.io import loadmat
>>> distance_distribution = loadmat('data/mean_distance_files/n01440764_distances.mat')
>>> print distance_distribution.keys()
['__header__', 'eucos', '__globals__', 'euclidean', 'cosine', '__version__']
>>> print distance_distribution['euclidean'].shape
(10, 1224)
Distances (e.g. euclidean or cosine or eucos) are computed from MAV to correctly classified training examples. In the above example, n01440764 category has 1300 training images. During training process, only 1224 images were classified correctly by the network. Thus, we build distance distribution using these correctly classified images. These distances are computed for each channel (for defn of channel see point 2.a). This process forms a distance distribution from the MAV. This distance distribution is used for Weibull fitting. In our experiments, we consider tail size for Weibull fitting as 20. This means, that 20 distances farthest away from MAV were used to estimate Weibull parameters. Again, since the distance computation process is very time comsuming, we are providing pre-computed distances for each class, so that you have access to the distance distribution from MAV for each category. It can be downloaded from
wget http://vast.uccs.edu/OSDN/data.tar
There are 1000 distance distributions, one for each category. The script used for computing distance distributions can be found in
python preprocessing/compute_distances.py
function weibull_tailfitting() in file evt_fitting.py . However, compute_openmax.py performs Weibull tailfitting for each category while computing probability for OpenMax
OpenMax probability for given image can be computed using following command.
python compute_openmax.py --image_arrname data/train_features/n01440764/n01440764_14280.JPEG.mat
The script accepts image feature files (features extracted from caffe as mentioned above). It computes openmax probability for the said image using default weibull tail sizes and other parameters. For more details, check the paper or get in touch with authors.
Fooling images are generously provided by Anh Nguyen and Prof. Jeff Clune from University of Wyoming. We will upload the fooling images and features extracted for fooling images in few days. When you use fooling images, please be sure to cite the following paper
@InProceedings{nguyen2015deep,
title={Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images},
author={Nguyen, Anh and Yosinski, Jason and Clune, Jeff},
booktitle={Computer Vision and Pattern Recognition (CVPR), 2015 IEEE Conference on},
year={2015},
organization={IEEE}
}