Robust Background Identification for Dynamic Video Editing

摘要

Extracting background features for estimating the camera path isrna key step in many video editing and enhancement applications.rnExisting approaches often fail on highly dynamic videos that arernshot by moving cameras and contain severe foreground occlusion.rnBased on existing theories, we present a new, practical method thatrncan reliably identify background features in complex video, leadingrnto accurate camera path estimation and background layering.rnOur approach contains a local motion analysis step and a globalrnoptimization step. We first divide the input video into overlappingrntemporal windows, and extract local motion clusters in each window.rnWe form a directed graph from these local clusters, and identifyrnbackground ones by finding a minimal path through the graphrnusing optimization. We show that our method significantly outperformsrnother alternatives, and can be directly used to improve commonrnvideo editing applications such as stabilization, compositingrnand background reconstruction.