We proposea high-throughput workflow tool-chain for georeferenced imageaugmentation that enables a significant increase in the amountof training samples. The presented pipeline exploits objects’segmentation masks to produce new realistic training scenesfrom target objects and various label-free backgrounds.
The augmnetation scheme and generated samples are presented below. Both target objects and backgrounds are from different images.
Ground truth, RGB image (a), models prediction without augmentation (b) and with object-based augmnetation (c) are shown below.
