KShapeNet: Riemannian network on Kendall shape space for Skeleton based Action Recognition

Deep Learning architectures, albeit successful in most computer vision tasks,\nwere designed for data with an underlying Euclidean structure, which is not\nusually fulfilled since pre-processed data may lie on a non-linear space. In\nthis paper, we propose a geometry aware deep learning approach for\nskeleton-based action recognition. Skeleton sequences are first modeled as\ntrajectories on Kendall's shape space and then mapped to the linear tangent\nspace. The resulting structured data are then fed to a deep learning\narchitecture, which includes a layer that optimizes over rigid and non rigid\ntransformations of the 3D skeletons, followed by a CNN-LSTM network. The\nassessment on two large scale skeleton datasets, namely NTU-RGB+D and NTU-RGB+D\n120, has proven that proposed approach outperforms existing geometric deep\nlearning methods and is competitive with respect to recently published\napproaches.\n