Learning to Denoise Raw Mobile UI Layouts for ImprovingDatasets at Scale

Gang Li, Gilles Baechler, Manuel Tragut, Yang Li

The layout of a mobile screen is a critical data source for UI designresearch and semantic understanding of the screen. However, UIlayouts in existing datasets are often noisy, have mismatches withtheir visual representation, or consists of generic or app-specifictypes that are difficult to analyze and model. In this paper, wepropose the CLAY pipeline that uses a deep learning approach fordenoising UI layouts, allowing us to automatically improve existingmobile UI layout datasets at scale. Our pipeline takes both thescreenshot and the raw UI layout, and annotates the raw layout byremoving incorrect nodes and assigning a semantically meaningfultype to each node. To experiment with our data-cleaning pipeline,we create the CLAY dataset of 59,555 human-annotated screenlayouts, based on screenshots and raw layouts from Rico, a publicmobile UI corpus. Our deep models achieve high accuracy withF1 scores of 82.7% for detecting layout objects that do not have avalid visual representation and 85.9% for recognizing object types,which significantly outperforms a heuristic baseline. Our work laysa foundation for creating large-scale high quality UI layout datasetsfor data-driven mobile UI research and reduces the need of manuallabeling efforts that are prohibitively expensive.

Knowledge Graph



Sign up or login to leave a comment