LISA Dataset for Classification

LISA is a dataset, mainly used for detection tasks. It contains images of traffic signs taken from video shots of driving vehicles. During my research, I came across Detecting Backdoor Attacks on Deep Neural Networks by Activation Clustering, which uses clustering as a defense against backdoor attacks. So in order to reproduce their results, I needed this dataset. So I thought I should make it available for other people in community whom may require the dataset as well. It is mainly tailored for use using Pytorch.

Usage

You could simply copy and paste lisa.py into your project and then use it same way you would use CIFAR10. For instance:

from lisa import LISA

dataset = LISA(root='path_to_data', download=True, train=True)

To reproduce the results of Activation Clustering paper, look at the activation_clustering_example.py code.

Dataset Properties

To train networks, one might need mean and std of the dataset. It's as following:

• mean: [0.4563, 0.4076, 0.3895]
• std: [0.2298, 0.2144, 0.2259]

Examples

There are some figures of the images presented in the examples folder. Here are the 5 classes that (I guess) authors of Activation Clustering paper use for classification:

Stop signs:

Yield Signs:

Warning Signs:

Speed Limit Signs:

Regulatory Signs: