Material for “ Shape Discovery from Unlabeled Image Collections ”

As described in the text of the main paper, this baseline is a sanity check to assure the difficulty of generating prototypical shapes. We manually partition the images into the “ideal” clusters, so that each cluster has 100% purity, and then simply average the aligned edge images, using the confidence weights given by the Pb detector [1]. The sanity check baseline helps to indicate the contribution made by our fragment weighting and prototype formation. Fig. 1 (a-c) show the prototypical shapes formed by the baseline on the Caltech images, ETHZ bounding box regions, and ETHZ expanded regions, respectively. For the Caltech images, the baseline clearly cannot discover the shape agreement, even though the input clusters were perfect. The baseline does pretty well to discover shape on the ETHZ bounding box regions, which is expected, since those regions are scalenormalized and aligned. The baseline performs worse on the ETHZ expanded regions due to clutter in the images; it discovers shapes with some accuracy for only a couple of categories (Applelogos and Bottles). (To compare against our method’s prototypical shapes, see Fig. 6 (b) (Caltech) and Fig. 7 (b,e) (ETHZ bounding box and expanded, respectively) in the main paper.) These results confirm that even with perfect clusters, simply stacking the edgemaps will not produce accurate shape models. Our method clearly outperforms this baseline for most of the generated shapes (a few are comparable), even without the advantage of starting with perfect clusters.