Detecting Duplicated Frames by Mapping Frames to 3D Skeletons

Frame Duplication is quite an effective approach to video tampering. In this paper, we propose a new method to detect such behavior. Inspired by the idea that skeletons capture the essential topologies of their corresponding objects, we use skeletons as a description of frames. We map superpixels to point cloud in 3D space and then extract the skeleton of the point cloud. By downsampling the resulted skeleton, for each frame we obtain a fixed-length feature, which comprises a topological component and a geometrical component. We use the topology-weighted location difference between skeletal nodes to measure the similarity between each pair of frames. We use precision, recall and F1 score to quantitatively evaluate the detection capability of the proposed method, and the experimental results confirm the efficacy of our method.