GeometryPaste: Geometry-Based Copy-Paste Data Augmentation for Instance Segmentation
Nicholas Dunn, Anurag Ghosh, and Christoph Mertz
RISS Working Papers Journal, 2023
Instance segmentation models require large datasets of annotated images to achieve adequate performance. However, annotated datasets are difficult to build or obtain. There are several large-scale datasets for common objects, but few exist for rare objects. Detecting rare objects has many practical applications, such as autonomous vehicles detecting roadwork objects. Copy-Paste is a data augmentation method for generating images and has been utilized successfully to improve instance segmentation performance. Prior works have studied both random Copy-Paste, where objects are randomly pasted onto images, and pasting objects based on the surrounding visual context. In this paper, we develop GeometryPaste, a method of pasting objects according to the geometry and context of the objects and background images. We build a small dataset of roadwork objects and fine-tune a pre-trained instance segmentation model to evaluate our method. Our results are compared against both baseline and random Copy-Paste APs. The results suggest that GeometryPaste may provide performance improvements over both baseline and random Copy-Paste augmentation for instance segmentation of rare object categories in small datasets.
