We present Spline-based Convolutional Neural Networks (SplineCNNs), a variantof deep neural networks for irregular structured and geometric input, e.g.,graphs or meshes. Our main contribution is a novel convolution operator basedon B-splines, that makes the computation time independent from the kernel sizedue to the local support property of the B-spline basis functions. As a result,we obtain a generalization of the traditional CNN convolution operator by usingcontinuous kernel functions parametrized by a fixed number of trainableweights. In contrast to related approaches that filter in the spectral domain,the proposed method aggregates features purely in the spatial domain. As a mainadvantage, SplineCNN allows entire end-to-end training of deep architectures,using only the geometric structure as input, instead of handcrafted featuredescriptors. For validation, we apply our method on tasks from the fields ofimage graph classification, shape correspondence and graph node classification,and show that it outperforms or pars state-of-the-art approaches while beingsignificantly faster and having favorable properties like domain-independence.