Learning-Based Trimap Generation for Video Matting

Kyoung-Rok Lee, Truong Q. Nguyen

Video Processing Lab

University of California, San Diego

Abstract

Object extraction is a critical operation for many content-based video applications. For these applications, a robust and precise extraction technique is required. This thesis proposes an efficient and accurate method for generating a trimap for video matting. We first segment the foreground using motion information and neighboring pixel coherence via graph cuts. Also, we estimate the parameters of a Gaussian Mixture Model for the foreground and background with segmented foreground and estimated static background. Next, we classify the pixels of each frame into models by performing maximum likelihood classification and generate a trimap which is an image consisting of three regions: foreground, background and unknown. Finally, we use the trimap as a guide in spectral matting for video matting. Our experimental results show that the proposed method yields accurate and natural object boundaries.

 

Results

1. Trimap generation and alpha matting results

(a) "aya dataset

Color Image Foreground Map Background Map Generated Trimap Alpha Matting Result

(a) "kyoung dataset

Color Image Foreground Map Background Map Generated Trimap Alpha Matting Result

2. Comparisons

Extraction Result Zoom-in

3. Composites

(a) "nana dataset

Color Image
Matte
Composite

(a) "natan dataset

Color Image
Matte
Composite