Nested Densely Atrous Spatial Pyramid Pooling and Deep Dense Short Connection for Skeleton Detection

Abstract

The skeleton shows the local symmetry and the shape/topology of the object, and it is utilized for human pose recognition, road detection, text detection, and the representation of industrial parts. However, the size of the skeleton is variable, which makes high-level feature representation difficult. Existing methods only attempt to integrate multilevel features but ignore the extraction of high-level features and multiscales of contextual information that are helpful for the skeleton detection task. Thus, the contributions of this article include two aspects. The first contribution is to propose a nested densely atrous spatial pyramid pooling method that connects the atrous convolutions with different dilation rates in a nested cascade mode, which can provide the multiscale denser contextual information, a larger receptive field, and more local features. The second one is to propose a deep dense short connection (DDSC) that explores the role of features at different levels in the task of skeleton detection. DDSC adopts concatenation to fuse high-level semantic information with shape information. The proposed method is evaluated on four common datasets, and the experimental results show the effectiveness of the proposed method.