Multiple Eigenspace Models for Scene Segmentation and Occlusion Removal

Abstract

We present a method that uses eigenspace models to segment an input image into a foreground and a background component. The algorithm segments the input image into the two components, removes the mutual occlusions among the objects of the foreground and background component, and reconstructs the occluded portions of both the foreground and the background component. The problem is formulated as a nonlinear optimization and an approximate solution is found by an iterative process that alternates between input image segmentation and component reconstruction, gradually improving the two components extracted from the input image. The novelty of this approach lies in the use of multiple eigenspaces to achieve a model-based segmentation of the input image in an iterative framework. This method yields segments that correspond to meaningful real-world objects even in the presence of occlusions, and these segments can be directly used for other tasks like object recognition. This method differs from the traditional segmentation algorithms as it is not obtained in the usual bottom-up manner but is model-guided. We demonstrate the utility of the algorithm in the segmentation and recognition of partially occluded humans in an office environment.