Thursday 23 January 2025
AI-generated videos are becoming increasingly prevalent, but detecting whether a video is real or fake can be a daunting task for humans. Researchers have created datasets to help train artificial intelligence (AI) models to identify these fake videos, but existing datasets have limitations.
A new dataset, called GenVidBench, aims to overcome these limitations by providing a comprehensive and diverse set of videos that AI models can use to learn how to detect AI-generated content. The dataset includes over 100,000 videos with rich semantic labels, making it the largest and most challenging dataset for AI-generated video detection.
The dataset is divided into two parts: cross-source and cross-generator tasks. In the former, AI models are trained on videos generated by different sources, such as text prompts or images, to test their ability to generalize across different types of content. In the latter, AI models are trained on videos generated by a single source but with varying levels of complexity, such as simple animations or complex scenes.
Researchers have used GenVidBench to train and evaluate various AI models, including SlowFast, I3D, F3Net, and others. The results show that existing AI models struggle to accurately detect AI-generated videos, especially when they are generated by different sources or with varying levels of complexity.
For example, the SlowFast model achieved an accuracy of only 41.66% on the GenVidBench dataset, while the I3D model achieved an accuracy of 49.23%. These results highlight the challenges faced by AI models in detecting AI-generated videos and underscore the need for more advanced and robust detection techniques.
The researchers behind GenVidBench hope that their dataset will help accelerate the development of more accurate AI video detection systems, which could have significant implications for a range of applications, from social media to law enforcement. By providing a comprehensive and challenging dataset, they aim to push the boundaries of what is possible in AI-generated video detection.
The dataset’s rich semantic labels also offer opportunities for researchers to explore new approaches to video analysis and understanding. For instance, by analyzing the semantic content of videos, researchers could develop more accurate methods for detecting AI-generated videos or identifying patterns in human behavior.
Overall, GenVidBench represents a significant step forward in the development of AI-generated video detection technology, offering a powerful tool for researchers and developers to improve their algorithms and ultimately enhance our ability to detect and analyze AI-generated content.
Cite this article: “GenVidBench: A Comprehensive Dataset for Detecting AI-Generated Videos”, The Science Archive, 2025.
Ai-Generated Videos, Video Detection, Dataset, Genvidbench, Machine Learning, Deep Learning, Image Recognition, Semantic Labels, Video Analysis, Video Understanding
