Visual Information Processing and Protection Group

VIPP Group datasets

Training and test data used in some of our papers

WILD (in-the-Wild Image Linkage Dataset): This dataset was created in collaboration with PoliMi, UniTn, La Sapienza, and UniCt. It consists of a closed set of 10,000 images from a mix of 10 commercial and open-source text-to-image generators, and an open set of another 10,000 images from 10 open-source generators of different nature. Thanks to an automatic prompt writing mechanism, the head-and-shoulder images are balanced with respect to sources and characteristics of the portrayed imaginary people. This dataset is a challenging benchmark which allows to benchmark synthetic image source attribution models in closed set attribution, open-set attribution, and robust attribution. The latter task is enabled by the post-processed replicas of test set and validation set images, and of half of the open set images. Each of these images has one copy with a single post-processing step, one with two steps, and one with three steps.

VippSent Dataset . In the field of Visual Sentiment Analysis (VSA), building image datasets labeled according to the emotions they evoke is essential, yet complex and prone to cognitive biases. Our dataset was developed through a scalable and customizable methodology, grounded in principles of semiotics and art theory, and is divided into three emotional classes: positive, negative, and neutral. The method combines common subjects (such as people, animals, and objects) with emotionally opposed adjective pairs and neutral adjectives identified through the semiotic square. It also includes artistic representations of emotional concepts, leveraging the ability of artists to express emotions and enhancing the dataset’s generalization capacity. This approach draws on the structures of visual language and the expressive power of art to support the classification of images based on the emotions they evoke, with a high degree of generalization. It is suitable for applications in areas such as marketing, opinion analysis, journalism, and the study of visual communication.

DeepStreets Dataset . Datasets used for training and testing detection for street deepfake.

Mobirise

UNISI & UNIFI Dataset. Collection of around 700 high quality TIFF uncompressed images (11 GB download)

A Universal Attack Against Histogram Based Image Forensics. Each folder contains host image, source image and the forgery before and after application of the counter-forensic scheme (7 hand-made splicings reported in the paper)

Universal Counterforensics of Multiple Compressed JPEG Images. Training and testing data used in Universal Counterforensics of Multiple Compressed JPEG Images, M. Barni, M. Fontani, and B. Tondi, IWDW 2014.

A Framework for Decision Fusion in Image Forensics based on Dempster-Shafer Theory of Evidence. Image dataset used in A Framework for Decision Fusion in Image Forensics based on Dempster-Shafer Theory of Evidence.

Identification of cut & paste tampering by means of double-JPEG detection and image segmentation. Image dataset used in Identification of cut & paste tampering by means of double-JPEG detection and image segmentation, ISCAS 2010.

© Copyright 2024 VIPP Group - All Rights Reserved

Mobirise.com