Feature Hourglass Network for Skeleton Detection

Geometric shape understanding provides an intuitive representation of object shapes. Skeleton is typical geometrical information. Lots of traditional approaches are developed for skeleton extraction and pruning, while it is still a new area to investigate deep learning for geometric shape understanding. In this paper, we build a fully convolutional network named Feature Hourglass Network (FHN) for skeleton detection. FHN uses rich features of a fully convolutional network by hierarchically integrating side-outputs with a deep-to-shallow manner to decrease the residual between the prediction result and the ground-truth. Experiment data shows that FHN achieves better performance compared with baseline on both Pixel SkelNetOn and Point SkelNetOn datasets.

1. Introduction

Geometric shape understanding provides an intuitive representation of object shapes, which can be used for foreground extraction, shape modeling, object proposal, et al. Even though deep learning approaches obtain great success for detection and segmentation tasks, it is still a new area to investigate deep learning for geometric shape understanding, especially for extracting topological and geometric information from shapes.