<h3>Overview</h3> <p><strong>Datasets introduced in</strong> "<em>Semi-Supervised Feature Embedding for Data Sanitization in Real-World Events</em>"</p> <p>It includes the links for the images that consist of the five datasets.</p> <ul> <li><strong>NotreDame:</strong> Remarkable partial destruction of the Parisian Cathedral by a fire in 2019.</li> <li><strong>Grenfell:</strong> 2017 tragic fire incident in the Grenfell Tower in London.</li> <li><strong>NationalMuseum:</strong> Total destruction of the National Museum in Brazil by flames in 2018.</li> <li><strong>BangladeshFire:</strong> Fast-moving fire in a district in Dhaka, that took place in 2019.</li> <li><strong>BostonMarathon:</strong> 2013 terrorist attack on the traditional Bostonian event.</li> </ul> <h3>Observations:</h3> <ul> <li>We did not publish the links for the positive images for the Grenfell dataset due to copyright reasons.</li> <li>BostonMarathon and Grenfell negative sets are pictures from the <em>Flickr100k</em> dataset.</li> <li>Since BostonMarathon positive samples contain frames from YouTube videos, we published the YouTube video URL and the frame number that was extracted.</li> <li>BostonMarathon positive sample contains augmentations, which are crops sized half of the original image. In this sense, columns <code>i</code> and <code>j</code> indicate the upper-left pixel of the crop. For example, if the image is <code>a × b</code>, the crop will be the square between the points: <code>(j, i)</code>, <code>(j + a/2, i)</code>, <code>(j, i + b/2)</code>, and <code>(j + a/2, i + b/2)</code>. </li> </ul> <h3>Media Content</h3> <p> Due to the terms of use from the social networks, we do not make publicly available the texts, images, and videos that were collected (only the links). However, we can provide some additional media content related to one (or more) events upon request by contacting the authors. </p>