A C3D-Based Convolutional Neural Network for Frame Dropping Detection in a Single Video Shot

Chengjiang Long; Eric Smith; Arslan Basharat; Anthony Hoogs; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017, pp. 86-94

Frame dropping is a type of video manipulation where consecutive frames are deleted to omit content from the original video. Automatically detecting dropped frames across a large archive of videos while maintaining a low false alarm rate is a challenging task in digital video forensics. We propose a new approach for forensic analysis by exploiting the local spatio-temporal relationships within a portion of a video to robustly detect frame removals. In this paper, we propose to adapt the Convolutional 3D Neural Network (C3D) for frame drop detection. In order to further suppress the errors due by the network, we produce a refined video-level confidence score and demonstrate that it is superior to the raw output scores from the network. We conduct experiments on two challenging video datasets containing rapid camera motion and zoom changes. The experimental results clearly demonstrate the efficacy of the proposed approach.