3D Shape Co-segmentation by Combining Sparse Representation with Extreme Learning Machine

Abstract

Unsupervised shape co-segmentation is proposed to segment a set of 3D shapes into meaningful parts without any labelled data. At the same time, a correspondence is created between the segmented parts. Usually, there are two main steps: correlation analysis and representation learning. In this paper, we propose an affinity matrix construction method based on parameter-free and high-efficiency simplex sparse representation to analysis correlation. This construction avoids the blindness of parameter setting. Based on the affinity matrix, we propose a co-segmentation approach via an unsupervised extreme learning machine to train a transform network for feature representation. This representation learning could attain good performance in lower embedding dimension. Therefore, co-segmentation can be implemented by clustering on lower dimensions embedding space. So the execution is more efficient. Moreover, once the transform network is trained, it can be applied to the data representation acquisition process without re-computing the affinity matrix. Experiments validate the method proposed in this paper. The method is unsupervised and can perform efficient and effective co-segmentation. Moreover, it also can deal with incremental co-segmentation when the data set is expanded.