Novelty is the quality of being different, new and unusual. Identifying it is an important issue in various fields such as anomaly detection in video. To detect the novelty, there are supervised learning methods that define and classify inliers and outliers, and unsupervised learning methods that define the distribution of inliers and identify whether objects are normal or abnormal. The former has limitations that the labeled data is required and the novelty which cannot be defined is not detected. To cope with the problems, the latter has recently been explored, but it is difficult to define an appropriate distribution for normal data and learn in an end-to-end manner due to unavailability of outliers. In this paper, we propose a novel one-class novelty detection method with constant curvature adversarial autoencoder. It consists of three components: an encoder, a decoder, and a discriminator. The encoder and discriminator interact with each other in adversarial and learn the distribution of normal data only. The decoder reconstructs the data to verify that the feature of the data is well extracted to the latent variable that is the output of the encoder. We also train the model to define a distribution for normal data as a constant curvature manifold, a non-Euclidean space, for the diversity of data distribution. The proposed method is verified with the well-known benchmark datasets: MNIST, CALTECH-256, and UCSD Pedestrian 1. For the area under curve as a measure of the performance, the proposed method shows the state-of-the-art performance with 0.87, 0.94, and 0.89 on average for the datasets, respectively.