Dataset corresponding to the paper "A privacy-preserving approach to identify riot-related footage on social media".

107,674 geolocated visual posts from a social media were collected during and after the 'Nahel Merzouk' riots in the summer 2023 in 7 French cities. These posts were fed to an image-to-text model (BLIP2-OPT-2.7B) to produce textual description of the visual content. This dataset contains those textual descriptions, along with the metadata (date, time, and location). A subset of the posts were also annotated as riot-related or not riot-related to train a BERT model. This subset is also provided in this database (see paper for more details).<br>Tables:<br>1. videos: Contains metadata about each video including location and timestamp information.2. captions: Contains all captions extracted from videos, with frame-level information.3. annotated_captions: Contains a subset of captions that have been manually annotated for riot-related content.4. annotated_videos: Contains manually annotated video-level labels for riot detection.5. split_annotated_videos: Defines the train/test split for annotated videos used in model training and evaluation.<br>