A graph cut based trimap propagation method for video matting
Year: 2017
Type: dissertation
Abstract: Alpha matting is an important task in image processing. It allows a user to extract foreground objects from natural images and compose them with different backgrounds, achieving subpixel-level precision. Due to its ill-posed nature, alpha matting requires additional inputs: for a single image, the user should specify the foreground, background, and unknown regions of the image using a trimap. However, when performing alpha matting on a video sequence, providing a trimap for each individual frame becomes tedious and time-consuming. We propose a novel method for propagating trimaps between frames of a video sequence. While most previous techniques require a binary segmentation for each frame to produce trimaps, our approach requires the user to define a trimap for only a few keyframes. To obtain trimaps for the remaining frames, we use the alpha channel (opacity map) computed for the previous frame to estimate the trimap for the next one, using a graph-cut-based approach. Results show that our algorithm is able to significantly reduce the amount of user effort for simple videos, while still resulting in considerable savings when applied to videos containing little temporal coherence.
Author: Marcos Henrique Backes
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2955912538