CAMION: Cascade Multi-input Multi-output Network for Skeleton Extraction

Sheng Fang, Kaiyu Li, Zhe Li

Abstract: Skeletonization is an important process of extracting the medial axis of the object shape while maintaining the original geometric and topological properties. Some recent studies have demonstrated that deep learning-based segmentation models can extract the main skeleton from objects more robustly. However, we find that the skeleton extracted by a vanilla segmentation process is always discontinuous and not accurate enough. In this paper, we propose a general cascade deep learning pipeline that achieves competitive performance only using a simple U-shape network. The semantic information contained in the shapes is limited, so we introduce a ConvNet with multi-source input and multi-task output, CAMION for short, on top of the basic shape-to-skeleton network. With the multi-source inputs, CAMION can converge faster than using only bi-nary shapes; and with the introduction of multi-task learning, relevant and suitable auxiliary tasks (e.g., feature point detection and contour extraction) bring considerable gains for the extraction of skeleton. Our code used in Pixel Skel-NetOn - CVPR 2022 challenge will be released at https://github.com/likyoo/CAMION-CVPRW2022.