Sparse Non-Negative Matrix Factorization for Mesh Segmentation

Abstract

In this paper, we present a method for 3D mesh segmentation based on sparse non-negative matrix factorization (NMF). Image analysis techniques based on NMF have been shown to decompose images into semantically meaningful local features. Since the features and coefficients are represented in terms of non-negative values, the features contribute to the resulting images in an intuitive, additive fashion. Like spectral mesh segmentation, our method relies on the construction of an affinity matrix which depends on the geometric properties of the mesh. We show that segmentation based on the NMF is simpler to implement, and can result in more meaningful segmentation results than spectral mesh segmentation.

SAC-NMF-Driven Graphical Feature Analysis and Applications

“…It is clear that GISIFs performs unstably on the low-quality model with obvious asymmetry, while our method achieves stable factorization results, which are visualized to be symmetric and part-aware, thus can help to detect symmetry when models are not provided in high quality. In order to further demonstrate the characteristics of our framework, we also make comparisons with the most related work of Mcgraw et al [17] and Liu et al [39] in segmentation and with the related works in saliency detection. The segmentation results of [17,39] are shown in (a) and (b), respectively, as can be seen from Figure 13.…”

“…In order to further demonstrate the characteristics of our framework, we also make comparisons with the most related work of Mcgraw et al [17] and Liu et al [39] in segmentation and with the related works in saliency detection. The segmentation results of [17,39] are shown in (a) and (b), respectively, as can be seen from Figure 13. Spectral mesh segmentation presented in [39] can achieve idea results for most models; however, it requires SVD decomposition process together with kmeans clustering, which are time consuming.…”

“…Spectral mesh segmentation presented in [39] can achieve idea results for most models; however, it requires SVD decomposition process together with kmeans clustering, which are time consuming. Sparse-NMF on affinity [17] can achieve ideal results when the segments' boundaries are distinct as the chair in Figure 13a, while it cannot promise good results on the others. Additionally, our method can achieve ideal segmentation results on different model due to the design of feature matrix and SAC-NMF.…”

“…Figure 13. Comparisons on segmentation, with (a-c) showing the results from [17,39], and our method.…”

“…Inspired by the Non-negative Matrix Factorization (NMF)-driven applications (esp. face recognition) in computer vision [13][14][15][16] and the related work to factorize the affinity matrix in graphics [17], we propose encoding the 3D model with part-aware bases that afford visually meaningful tools for further applications.…”