Abstract. A method for exploiting the information in low-level image segmentations for the purpose of object recognition is presented. The key idea is to use a whole ensemble of segmentations per image, computed on different random samples of image sites. Along the boundaries of those segmentations that are stable under the sampling process we extract strings of vectors that contain local image descriptors like shape, texture and intensities. Pairs of such strings are aligned, and based on the alignment scores a mixture model is trained which divides the segments in an image into fore- and background. Given such candidate foreground segments, we show that it is possible to build a state-of-the-art object recognition system that exhibits excellent performance on a standard benchmark database. This result shows that despite the inherent problems of low-level image segmentation in poor data conditions, segmentation can indeed be a valuable tool for object recognition in real-world images. 1

Abstract. The complexity of visual representations is substantially limited by the compositional nature of our visual world which, therefore, renders learning structured object models feasible. During recognition, such structured models might however be disadvantageous, especially under the high computational demands of video. This contribution presents a compositional approach to video analysis that demonstrates the value of compositionality for both, learning of structured object models and recognition in near real-time. We unite category-level, multi-class object recognition, segmentation, and tracking in the same probabilistic graphical model. A model selection strategy is pursued to facilitate recognition and tracking of multiple objects that appear simultaneously in a video. Object models are learned from videos with heavy clutter and camera motion where only an overall category label for a training video is provided, but no hand-segmentation or localization of objects is required. For evaluation purposes a video categorization database is assembled and