Consistent 3D human body segmentation based on combinatorial descriptor in spectral domain

Haoyang Xie, Yueqi Zhong

Abstract: Consistent 3D human body segmentation plays a vital role in many human-oriented applications. Recently research used supervised methods to achieve state-of-the-art performance. However, requiring massive labelled data and tedious training is a costly process. Unsupervised methods do not need labelling and training but struggle to achieve consistent segmentation for a non-rigid deformable mesh. Moreover, the segmentation style is also fixed. In the paper, we aim to achieve high-performance, consistent 3D human mesh segmentation avoiding fully-labelled data and time-consuming training. Specifically, this paper designs a Laplacian operator by incorporating mesh saliency, in which a face-level filter is proposed to improve the detection of concave vertices. Accordingly, we construct a combinatorial descriptor by explicitly employing the global and local attributes derived from the spectrum of the proposed saliency Laplacian operator to achieve consistent segmentation in the spectral domain. An automatic determination mechanism is adopted to determine the number of segments. Extensive experimental results demonstrate that the presented method is effective and efficient for many 3D meshes, especially for the human body shape. The segmentation results are comparable to other state-of-the-art performances without requiring time-consuming labelling and training on large-scale datasets.