A three-level benchmark dataset for spatial and temporal forensic analysis of videos

A three-level benchmark dataset for spatial and temporal forensic analysis of videos

연구 분야: Analysis

학회: Machine Vision and Applications

Video Forensics keeps developing new technologies to verify the authenticity of digital videos. The existing datasets have limitations including unrealistic and poor-quality tampering, small size, few types of forgeries, lack of range of video content, different lighting conditions, and a range of camera models. This paper proposes the COMSATS Structured Video Tampering Evaluation Dataset (CSVTED), a three-level benchmark dataset organized by tampering quality and video complexity. This dataset includes a diversity of tampering in the spatial and temporal domains such as frame duplication, deletion, insertion, copy-move, and splicing. The dataset aims to facilitate the evaluation of video forgery detection methods by providing 1047 videos (133 original and 914 tampered), captured by multiple cameras in different lighting conditions (morning, noon, evening, night, fog). To develop the benchmark dataset, videos are tampered with a variable number of duplicated/deleted/inserted frames as well as Event-Object-Person (EOP) based tampering. Special care has been taken to ensure minimal abrupt changes in tampered videos by using Structural Similarity Index Measure (SSIM) and Optical Flow (OF) to determine the optimal positions for duplication/insertion/deletion in the video. Taking into account the direction of motion of objects in the video, these techniques aid in seamlessly integrating the tampered frames while maintaining visual coherence. Furthermore, the videos in CSVTED depict natural scenes after the tampering process and are in the common formats of avi, mp4, or mov. This dataset will be publicly available for researchers in the domain of video forensic analysis.