We address the problem of estimating a scene's background, given a set
of non-time sequence photographs captured from the same view point and
under similar illumination conditions. One example application would
be to capture a clean, unoccluded shot of a crowded public place. We
also demonstrate the application of background estimation for the
removal of ghosting artifacts during the construction of high dynamic
range images from multiple exposures. Our approach consists of
defining the scene's background as a labeling over the set of input
images, assigning each labeling a cost, and minimizing the associated
cost function.The cost function penalizes deviations from the
following model assumptions: background objects are more likely to
appear across the input images, and background objects are stationary.
We approximate object likelihood using an entropy weighted probability
estimate, and object stationariness using a motion boundary
consistency term. We furthermore introduce a constraint that prevents
the algorithm from classifying fractions of a transient object as
background, i.e.\ to cut through objects. Our method benefits from the
appropriate selection of color spaces for each task. The cost
function is minimized by a graph cut based expansion move algorithm,
and the final result is composed using Poisson blending. Our
contribution is the definition of an automatic method for consistent
background estimation, and its application to ghost removal in high
dynamic range image reconstruction.