Data augmentation is a key element in training high-dimensional models. In this approach, one synthesizes new observations by applying pre-specified transformations to the original training data; e.g. new images are formed by rotating old ones. Current augmentation schemes, however, rely on manual specification of the applied transformations, making data augmentation an implicit form of feature engineering. With an eye towards true end-to-end learning, we suggest learning the applied transformations on a per-class basis. Particularly, we align image pairs within each class under the assumption that the spatial transformation between images belongs to a large class of diffeomorphisms. We then learn a class-specific probabilistic generative models of the transformations in a Riemannian submanifold of the Lie group of diffeomorphisms. We demonstrate significant performance improvements in training deep neural nets over manually-specified augmentation schemes. Our code and augmented datasets are available online.

We present an autoencoder that leverages learned representations to better measure similarities in data space. By combining a variational autoencoder (VAE) with a generative adversarial network (GAN) we can use learned feature representations in the GAN discriminator as basis for the VAE reconstruction objective. Thereby, we replace element-wise errors with feature-wise errors to better capture the data distribution while offering invariance towards e.g. translation. We apply our method to images of faces and show that it outperforms VAEs with element-wise similarity measures in terms of visual fidelity. Moreover, we show that the method learns an embedding in which high-level abstract visual features (e.g. wearing glasses) can be modified using simple arithmetic.

We present a new method for automatic classification of indirect immunoflourescence images of HEp-2 cells into different staining pattern classes. Our method is based on a new texture measure called shape index histograms that captures second-order image structure at multiple scales. Moreover, we introduce a spatial decomposition scheme which is radially symmetric and suitable for cell images. The spatial decomposition is performed using donut-shaped pooling regions of varying sizes when gathering histogram contributions. We evaluate our method using both the ICIP 2013 and the ICPR 2012 competition datasets. Our results show that shape index histograms are superior to other popular texture descriptors for HEp-2 cell classification. Moreover, when comparing to other automated systems for HEp-2 cell classification we show that shape index histograms are very competitive; especially considering the relatively low complexity of the method.