Curriculum Domain Generalization For Computer Vision

Domain generalization aims to train models to perform well on unseen domainswithout access to data from those domains during training. ADRMX (AdditiveDisentanglement of Domain Features with Remix Loss) is an augmentation baseddesign to improve generalization to unseen domains. ADMRX disentangles domaininvariantand domain-specific features via an additive architecture and applies alatent-space remix loss, mixing same-class representations across source domains togenerate synthetic samples. Building on the ADRMX method, which mixes featurerepresentations of same-class samples across different domains, this thesis introducesProgressive Feature Alignment (PFA). PFA is a curriculum-driven remixing strategy.Remixing proceeds from the closest to the most distant pairs of domains, withmixing coefficients dynamically adjusted based on class centroid distances across domainsto prevent unrealistic blending of dissimilar features and reduce noise in theresulting synthetic examples. By organizing feature remixing according to semanticproximity, PFA enables a gradual adaptation to increasingly challenging shifts.Under the leave-one-domain-out protocol on the PACS and OfficeHome benchmarks,PFA consistently outperforms ADRMX and other state-of-the-art domain generalizationtechniques, yielding especially strong gains on the more challenging Office-Home dataset. These results demonstrate that a curriculum-driven approach tofeature remixing can substantially enhance the robustness of computer vision modelsto complex domain variation, suggesting new directions for tackling severe shiftsin unseen data. The implementation of my method is available at: https://github.com/SormehSerp/PFA_