oa Spider: A System for Finding Illegal 3D Video Copies

Three-dimensional (3D) videos are getting quite popular, and equipment for recording and processing them are becoming affordable. Creating 3D videos is expensive. Thus, protecting 3D videos against illegal copying is an important problem. We present a novel system for finding 3D video copies. Our system also identifies the location of the copied part in the reference video. The system can be used, for example, by video content owners, video hosting sites, and third-party companies to find illegally copied 3D videos. To the best of our knowledge, this is the first complete 3D video copy detection system in the literature.

Detecting 3D video copies is a challenging problem. First, comparing numerous numbers of frames from potential copies against reference videos is computationally intensive. Second, many modifications occur on copied videos; some of them are intentional to avoid detection and others are side effects of the copying process. For example, a copied video can be scaled, rotated, cropped, transcoded to a lower bit rate, or embedded into another video. The contrast, brightness, and colors of a video can also be manipulated. Furthermore, 3D videos come in various encoding formats, including stereo, multiview, video plus depth, and multiview plus depth. Changing the format is possible during copying, which complicates the detection process. Finally, new views can be synthesized from existing ones. These views display the scene from different angles, and thus reveal different information than in original views. For example, an object occluded in one view could appear in another.

We implemented the proposed system and evaluated its performance using many 3D videos. We created a large set of query videos with 284 videos to represent all practical scenarios. Our results show that the proposed system achieves high precision and recall values in all scenarios. Specifically, our system results in 100% precision and recall when copied videos are unmodified parts of original videos, and it produces more than 90% precision and recall when copied videos are subjected to various transformations. Even in extreme cases where each video is subjected to five different transformations, our system yields more than 75% precision and recall.