Skeletonization in natural images and its application to object recognition

Object skeletons are utilized to represent objects because they clarify the structural relationship between various parts of the object. Skeletonization in natural images is a challenging problem since it is necessary for the extractor to capture both contextual and local information. These types of information then must be utilized to determine the scale of every individual skeleton pixel. To handle this challenge, we develop a fully convolutional network with multiple scale-associated side outputs. We introduce a scale-associated side output for every stage based on the relationship between the receptive field sizes of the sequential stages in the network and the skeleton scales they can capture. We supervise each stage by guiding the scale-associated side outputs toward the groundtruth skeletons with varying scales. We then fuse the responses of multiple scale-associated side outputs in a scale specific way, and eventually we can effectively localize skeleton pixels with multiple scales. Our method performs preferably on two skeletonization datasets and significantly outperforms other competitors. Additionally, the usefulness of the obtained skeletons is verified on extensive object recognition applications, including symmetric part segmentation, object proposal detection, road detection, and text line proposal generation.