"Sub-element_Retrieval"

"The dataset is a publicly available benchmark designed for fine-grained sub-element retrieval in printed fabrics, a task in which an isolated local motif is used to retrieve complete images containing the same or a semantically consistent target. The dataset addresses the shortage of specialized resources for this local-to-global retrieval setting and consists of two complementary parts: a controllable simulated dataset, SimulatedPrintedFabrics-17k, and a real-world validation dataset, RealPrintedFabrics-13k. The simulated portion enables supervised training and quantitative evaluation, while the real portion supports practical assessment under realistic textile conditions. The simulated dataset is generated by combining collected sub-elements with diverse fabric-like backgrounds and arranging them through multiple layout strategies, including regular tiling, randomized placement, and transparency-based blending. Random cropping is further applied to increase variation in visible extent, composition, and background proportion. Because each synthesized image is created from known source elements, explicit query\u2013gallery relevance labels can be defined reliably, making the dataset suitable for controlled retrieval experiments. In contrast, the real dataset contains more challenging printed fabrics with richer textures, denser motif repetition, more diverse decorative combinations, and stronger appearance variation. Together, these two subsets provide a useful simulated-to-real benchmark for developing and evaluating retrieval methods in complex visual scenes."