scotopic

Train and test scotopic classifiers on MNIST and CIFAR (c) 2016 Bo Chen

About

This codebase corresponds to the following publication:

[1] Bo Chen and Pietro Perona, Seeing into Darkness: Scotopic Visual Recognition, ArXiv 2016

• Train and test WaldNet, photoic classifier, ensemble classifier, and WaldNet with dynamic light level estimation.

• All models are tested against different camera noises.

• Includes plotting code to reproduce figures in [1].

Installation

• This codebase depends on and modifies MatConvNet (beta18). The MatConvNet is included and must be installed as explained here. Use different compile options depending on whether your machine has GPUs.

• In Matlab, go to the scotopic directory and type

addpath(genpath(pwd))

• Modify getScotopicConfig.m to have data_path point to your custom data directory

Training scotopic classifiers and evaluate on MNIST and CIFAR

• Now you can train and test scotopic classifiers. To start, run

cnn_scotopic_mnist_demo(1)

to train a WaldNet on the MNIST dataset. The argument to the function indicates what type of scotopic model you are training: 1: WaldNet; 2: Rate classifier; 3-6: Specialists at PPP=[.22, 2.2, 22, 220], respectively; 7: Ensemble; and 8: WaldNet with light level estimation. Each model takes from 20 minutes to a day to train depending on your workstation's config.

• After you have trained all of them (training in order is recommended), visualize the results by running

plot_SAT_mnist(FIGURE_FOLDER)

where FIGURE_FOLDER should be a folder you created in advance to store the result figures. You should be able to regenerate figures in [1] this way.

• The results from CIFAR10 may be obtained par simile.

Testing scotopic classifiers under camera noises

After all the models are trained, run

cnn_scotopic_noise(DATASET)

where DATASET is either "mnist" or "cifar". The evaluation should take about four hours if your machine support Matlab's parpool environment.

After that's finished you may visualize the robustness analysis by running

plot_sensor_noise_effect(FIGURE_FOLDER)

where FIGURE_FOLDER should be a folder you created in advance to store the result figures.

Contact

Questions please email [email protected]