Depth-Of-Field-Based Alpha-Matte Extraction

In compositing applications, objects depicted in images frequently have to be separated from their background, so that they can be placed in a new environment. Alpha mattes are important tools aiding the selection of objects, but cannot normally be created in a fully automatic way. We present an algorithm that requires as input two images - one where the object is in focus, and one where the background is in focus - and then automatically produces an alpha matte indicating which pixels belong to the object. This algorithm is inspired by human visual processing and involves nonlinear response compression, center-surround mechanisms as well as a filling-in stage. The output can then be refined with standard computer vision techniques. Copyright © 2005 by the Association for Computing Machinery, Inc