[Paper-Project] Video Copy Detection
Topic: Video Matching
Keywords: Video copy detection, Deep learning, AI
Motivation: The booming of GenAI techniques has been opening up a new horizon for content creation. With the help of GenAI applications, text, image, video music are created faster with high quality than ever. Our project is to tackle a specific problem in video matching. Given a short clip (a news clip registered by a journalist for instance), we are interested in deploying an Information Retrieval Engine to find relevant, similar videos in given video database (of a news broadcasting company). These resulting videos could be used for creating official news videos related to the content of the inputted clip.
Tasks: Input: Dataset from MetaAI
In this project, we will do two phases:
- First: Check if a video is copy or not. In this task, we will extract video into frames, then embedding these frames used CLIP model. After that, to reduce the amount of calculation cost, we will check if a video is edited or not because we see that if a video is edited -> it is usually a copied video by let the embeddings into a Roberta Model. After checking a video is edited, or not. If unedited -> convert it to a random vector. If edited -> extract by swinv2 model -> We can learn the embeddings by contrastive learning. In inference phase, we extract both query videos and reference videos into embedding -> We use KNN to get the top equivalent features.
**- Second: If the video is copied -> what segment? ( Updating ) ** link infer: https://drive.google.com/drive/folders/1SGnS0WQhCcx-Ut39CUvWZVuylY8Ku92u?usp=sharing **