Learning Uniform Hyperspherical Centers for Open and Closed Set Recognition

Open-set recognition remains a challenging problem, particularly when traditional closed-set classifiers are unable to generalize to unseen classes. In this paper, we propose a unified framework that leverages hyperspherical embeddings with learnable class centers for both open-set and closed-set recognition. Each class is represented by a center point uniformly distributed on the surface of a hypersphere, and training samples are encouraged to form compact clusters around their respective centers. Unlike previous methods that constrain features to the hypersphere boundary, we adopt a Euclidean distance-based formulation to improve flexibility and generalization. Our approach jointly optimizes class centers and feature representations, eliminating the reliance on predefined center locations. Additionally, we introduce a mechanism to incorporate background or unknown samples during training to further enhance open-set robustness. Extensive experiments on multiple benchmarks demonstrate that our method outperforms existing approaches, achieving state-of-the-art accuracy in both open-set and closed-set settings. The source code for the proposed approach is available at https://github.com/Cevikalp/dudch