Interactive Generation of Image Variations for Copy-Paste Data Augmentation

CHI 2023 LBW

By : Keita Higuchi, Taiyo Mizuhashi, Fabrice Matulic, Takeo Igarashi

In machine learning, data augmentation is an important technique to artificially increase the amount of training data by generating variations, e.g., geometric and colour transformations. Simple data augmentation such as scaling and rotation is already provided by existing tools, but advanced data augmentation such as copy-paste image composition requires coding. Such composition operations are difficult to intuitively define in coding environments as typically there is no visual confirmation of generated images. Therefore, composition-based augmentations are not frequently used by developers. To address this issue, we propose a dedicated graphical tool. Contrary to image operations of standard graphics editors designed to produce a single image, our tool creates multiple image variations to be used as training data. The editor allows the user to visually and interactively set parameter ranges for transformations, and quickly review synthesized images based on the parameters. We report performance evaluations and user studies with machine learning experts.

  • Twitter
  • Facebook