2020/04/02 by Maxime Bucher, Bucher, Maxime, Tuan-Hung Vu +5
Biochemistry, Genetics and Molecular Biology · Computer Science · Medicine · #COVID-19 diagnosis using AI #Cancer-related molecular mechanisms research #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences
paper · pdf · doi:10.48550/arxiv.2004.01130
openalex publication_date 2020/04/02 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
In this work, we define and address a novel domain adaptation (DA) problem in\nsemantic scene segmentation, where the target domain not only exhibits a data\ndistribution shift w.r.t. the source domain, but also includes novel classes\nthat do not exist in the latter. Different to "open-set" and "universal domain\nadaptation", which both regard all objects from new classes as "unknown", we\naim at explicit test-time prediction for these new classes. To reach this goal,\nwe propose a framework that leverages domain adaptation and zero-shot learning\ntechniques to enable "boundless" adaptation in the target domain. It relies on\na novel architecture, along with a dedicated learning scheme, to bridge the\nsource-target domain gap while learning how to map new classes' labels to\nrelevant visual representations. The performance is further improved using\nself-training on target-domain pseudo-labels. For validation, we consider\ndifferent domain adaptation set-ups, namely synthetic-2-real, country-2-country\nand dataset-2-dataset. Our framework outperforms the baselines by significant\nmargins, setting competitive standards on all benchmarks for the new task. Code\nand models are available at https://github.com/valeoai/buda.\n