Video forgery detection and localization based on 3D patchmatch

A new algorithm is proposed for the detection and accurate localization of copy-move video forgeries. Major ingredients of the proposed algorithm are i) noise-resilient rotation-invariant features, ii) dense-field matching over the whole video by means of a suitably modified fast algorithm for approximate nearest neighbor search, iii) ad hoc fast post-processing for forgery detection and false alarm removal. Experiments carried out on a dataset publicly available for this task show promising results, suggesting also a number of directions for future research.