Key-frame Based Spatiotemporal Scribble Propagation
We present a practical, key-frame based scribble propagation framework.
May 4, 2015
Pelin Dogan (Disney Research/ETH Joint M.Sc.)
Tunc Aydin (Disney Research)
Nikolce Stefanoski (Disney Research)
Aljoscha Smolic (Disney Research)
We present a practical, key-frame based scribble propagation framework. Our method builds upon recent advances in spatiotemporal filtering by adding key-components required for achieving seamless temporal propagation. To that end, we propose a temporal propagation scheme for eliminating holes in regions where no motion path reaches reliably. Additionally, to facilitate the practical use of our technique, we formulate a pair of image edge metrics influenced from the body of work on edge-aware filtering and introduce the “hybrid scribble propagation” concept where each scribble’s propagation can be controlled by user defined edge stopping criteria. Our method improves the current state-of-the-art in the quality of propagation results and in terms of memory complexity. Importantly, our method operates on a limited, user-defined temporal window and therefore has a constant memory complexity (instead of linear) and thus scales to arbitrary length videos. The quality of our propagation results is demonstrated for various video processing applications such as mixed HDR video tone mapping, artificial depth of field for video and local video recoloring.