Part-based deformable object detection with a single sketch

Object Detection using shape is interesting since it is well known that humans can recognise an object simply from its shape. Thus, shape-based methods have great promise to handle a large amount of shape variation using a com- pact representation. In this paper, we present a new algorithm for object detection that uses a single reasonably good sketch as a reference to build a model for the object. The method hierarchically segments a given sketch into parts using an automatic algorithm and estimates a di erent a ne transfor- mation for each part while matching. A Hough-style voting scheme collects evidence for the object from the leaves to the root in the part decomposi- tion tree for robust detection. Missing edge segments, clutter and generic object deformations are handled by exibly following the contour paths in the edge image that resemble the model contours. E cient data-structures and a two-stage matching approach assist in yielding an e cient and robust system. Results on ETHZ and several other popular image datasets yield promising results compared to the state-of-the-art. A new dataset of real-life hand-drawn sketches for all the object categories in the ETHZ dataset is also used for evaluation.Accepted versio