Learning-based Sampling for Natural Image Matting

We present a new sampling-based natural matting tech- nique that utilizes a pair of novel sampling networks for estimating background and foreground colors of pixels in unknown image regions.

We present a new sampling-based natural matting tech- nique that utilizes a pair of novel sampling networks for estimating background and foreground colors of pixels in unknown image regions. These color predictions are then provided as additional input to a separate matting net- work, along with an input image and trimap. We show that our data-driven approach is advantageous over previous sampling-based matting approaches that rely on explicitly defined rules for sample gathering and selection. In an ab- lation study, we present evidence that sampling networks consistently improve matting quality even when plugged into different matting network architectures. Our proto- type implementation of the proposed technique achieves top ranking in current alpha matting benchmarks.