ForgeryNet is a large-scale benchmark dataset for digital forgery analysis, constructed by Yinan He, Bei Gan, Siyu Chen, Yichun Zhou, Guojun Yin, Luchuan Song, Lu Sheng, Jing Shao, Ziwei Liu, and others. Its background is that the rapid development of realistic synthesis technology has increasingly blurred the boundary between real and manipulated images, making benchmarking and advancing digital forgery analysis an urgent issue, while existing face forgery datasets either have limited diversity or only support coarse-grained analysis.
In terms of use, ForgeryNet is mainly used for benchmarking and research on face forgery detection and analysis.
In terms of main capabilities, the dataset has a scale of 2.9 million images and 221,247 videos, including manipulations from 7 image-level methods and 8 video-level methods from around the world. Each video and image is associated with its label. The dataset has unified annotations across image- and video-level data, spanning four tasks, of which those explicitly mentioned include: 1) Image forgery classification, covering binary (real/fake), three-way (real/fake, fake with identity replacement forgery method/fake with identity preservation forgery method) and n-way (real and 15 respective forgery methods) classification; 2) Spatial forgery localization, segmenting the manipulated regions of forged images with their corresponding source real images.
In terms of usage, the dataset is available for download. The page lists multiple files: Training.tar, Validation.tar, public_test_images.tar, public_test_videos.tar, along with MD5 checksums. After downloading the training set, use the command cat Training.tar.* >> Training.tar to merge. The label list can be downloaded by clicking "Download * List", and the dataset can be downloaded by clicking "Download * Set". If Google Drive is accessible, other download methods are also provided.
In terms of applicable scenarios, it is suitable for academic research and benchmark evaluation of face forgery detection and analysis that require large-scale, multi-task, uniformly annotated data.
In terms of limitations, the official materials do not provide further explanation regarding fees, pricing, platform support, specific usage steps, or performance metrics, and none is currently provided.
