While mid-level perceptual cues have long been of interest in the human vision community, their role in computer vision has remained limited. In this report, we evaluate several algorithms which make use of mid-level processing in order to improve boundary detection. Our first technique builds a probabilistic model of the relation between prototypical local shapes of edges and the presence or absence of a boundary. We also present a more explicit local model of curvilinear continuity using piecewise linear representations of contours and the Constrained Delaunay Triangulation (CDT). Lastly we consider a global random field on the whole CDT which captures continuity along with the frequency of different of junction types. All three models are trained on human labeled groundtruth. We measure how each model, by incorporating mid-level structure, improves boundary detection. To our knowledge, this is the first time that such cues have been shown quantitatively useful for a large set of natural images. Better boundary detection has immediate application in the problem of object recognition.