3D Shapes Classification Using Intermediate Parts Representation

Jan Hula, David Mojzísek, David Adamczyk

Abstract: We describe a novel approach for 3D shape classification which classifies the shape based on a graph of its parts. To segment out the parts of a given object, we train a shape segmentation network to mimic the segments obtained from an offline co-segmentation method. Using the predicted segments, our approach constructs a spatial graph of the parts which reflects the spatial relations between them. The graph of parts is finally classified by a Tensor Field Network - a type of a graph neural network which is designed to be equivariant to rotations and translations. Therefore, the classification of the spatial graph of parts is not influenced by the choice of the coordinate frame. We also introduce a data augmentation method which is particularly suitable to our setting. A preliminary experimental results show that our method is competitive with the standard approach which does not detect parts as an intermediate step. The intermediate representation of parts makes the whole model more interpretable.