In this paper, we present a new segmentation-based 2D mosaic framework. Most of current mosaic algorithms do not explicitly remove moving objects from images before registration, so that they often fail when the size of the moving objects is relatively large. To solve this problem, we first segment moving objects from the input images using the tensor voting framework, and then only the remaining backgrounds are processed for the background mosaic. The second mosaicking step is straightforward because the first motion segmentation step also produces very accurate dense matches. By providing comparative examples, we show that the quality of the background mosaics can be significantly improved by our framework.