View article

We propose a method to automatically extract a sketch of a common object structure present in a small set of real world weakly-labeled images. Applying a part-based deformable contour matching technique gives the location of repeatable contours. An initial deformable search strategy selects a set of salient, repeatable contours robust to a large range of non-rigid deformations. A contour completion technique based on a locally greedy bi-directional search strategy is adopted to merge the repeatable contour fragments for obtaining a complete shape. The output of our algorithm is used as an input to a sketch-based object-recognizer with results that are either better, or on par with those obtained with the ground truth sketches provided with the dataset.