Boundary localization in an image pyramid

Abstract A pyramid is a stack of successively smaller versions of a given image, with size and resolution decreasing exponentially. Large objects can be detected on high levels of the pyramid at low cost, since the images at high levels are small, but the boundaries of objects detected in this way are not accurately localized. In principle, we can “project” a boundary down to the next lower level of the pyramid; use local search at that level to localize it more precisely; and repeat the process at successively lower levels. In practice, however, it is not clear what cost function should be used in defining the “optimal” boundary at each level. We show that with proper choice of criterion, the boundaries obtained in the full-resolution image (at the base of the pyramid) seem very reasonable perceptually.