Extracting object skeleton using a unique Fully Convolutional Network. We build a new dataset, named sk506 (download) for object skeleton detection in natural images. Our insight focuses on the relationship between receptive fields and skeleton scales, resulting in outperforming existing methods on mainstream public datasets.
The research paper “Object Skeleton Extraction in Natural Images by Fusing Scale-associated Deep Side Outputs” (pdf) has been accepted by CVPR 2016. The proposed approach achieves the state-of-the-art performance at two main datasets including the noval sk506 dataset. And we share all codes at github (code), managed by Kai Zhao. Just try it!