Batch normalization embeddings for deep domain generalization

• Published in: Pattern recognition Vol. 135; p. 109115
• Main Authors: Segu, Mattia, Tonioni, Alessio, Tombari, Federico
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2023
• Subjects: Domain generalization; Domain representation learning; Learning from multiple sources; Domain representation learning; Learning from multiple sources; Domain generalization
• ISSN: 0031-3203; 1873-5142
• DOI: 10.1016/j.patcog.2022.109115

•We propose to accumulate domain-specific batch normalization statistics accumulated on convolutional layers to map image samples into a latent space where membership to a domain can be measured according to a distance from domain batch population statistics•We propose to use this concept to learn a lightweight ensemble model that shares all parameters excepts the normalization statistics and can generalize better to unseen domains•Compared to previous work, we do not discard domain-specific attributes but exploit them to learn a domain latent space and map unknown domains with respect to known ones•We show a significant increase in classification accuracy over current state-of-the-art techniques on popular domain generalization benchmarks: PACS, Office-31 and Office-Caltech. [Display omitted] Domain generalization aims at training machine learning models to perform robustly across different and unseen domains. Several methods train models from multiple datasets to extract domain-invariant features, hoping to generalize to unseen domains. Instead, first we explicitly train domain-dependent representations leveraging ad-hoc batch normalization layers to collect independent domain’s statistics. Then, we propose to use these statistics to map domains in a shared latent space, where membership to a domain is measured by means of a distance function. At test time, we project samples from an unknown domain into the same space and infer properties of their domain as a linear combination of the known ones. We apply the same mapping strategy at training and test time, learning both a latent representation and a powerful but lightweight ensemble model. We show a significant increase in classification accuracy over current state-of-the-art techniques on popular domain generalization benchmarks: PACS, Office-31 and Office-Caltech.