3D Mesh Segmentation via Multi-branch 1D Convolutional Neural Networks
share
3D mesh segmentation is an important research area in computer graphics, and there is an increasing interest in applying deep learning to this challenging area. We observe that 1) existing techniques are either slow to train or sensitive to feature resizing and sampling, 2) in the literature there are minimal comparative studies and 3) techniques often suffer from reproducibility issue. These hinder the research development of supervised segmentation tasks. This study contributes in two ways. First, we propose a novel convolutional neural network technique for mesh segmentation, using 1D data and filters, and a multi-branch network for separate training of features of three different scales. We also propose a novel way of computing conformal factor, which is less sensitive to small areas of large curvatures, and improve graph-cut refinement with the addition of a geometric feature term. The technique gives better results than the state of the art. Secondly, we provide a comprehensive study and implementations of several deep learning techniques, namely, neural networks (NNs), autoencoders (AEs) and convolutional neural networks (CNNs), which use an architecture of at least two layers deep. The significance of the study is that it offers a novel fast and accurate CNN technique, and a comparison of several other deep learning techniques for comparison.
READ FULL TEXT