In this paper, we study the problem of semantic annotation on 3D models thatare represented as shape graphs. A functional view is taken to representlocalized information on graphs, so that annotations such as part segment orkeypoint are nothing but 0-1 indicator vertex functions. Compared with imagesthat are 2D grids, shape graphs are irregular and non-isomorphic datastructures. To enable the prediction of vertex functions on them byconvolutional neural networks, we resort to spectral CNN method that enablesweight sharing by parameterizing kernels in the spectral domain spanned bygraph laplacian eigenbases. Under this setting, our network, named SyncSpecCNN,strive to overcome two key challenges: how to share coefficients and conductmulti-scale analysis in different parts of the graph for a single shape, andhow to share information across related but different shapes that may berepresented by very different graphs. Towards these goals, we introduce aspectral parameterization of dilated convolutional kernels and a spectraltransformer network. Experimentally we tested our SyncSpecCNN on various tasks,including 3D shape part segmentation and 3D keypoint prediction.State-of-the-art performance has been achieved on all benchmark datasets.
translated by 谷歌翻译