Semi-Self-Supervised Segmentation of Oranges with Small Sample Sizes

Abstract. Visually detecting and masking fruit can be regarded as a vital first step in autonomous harvesting. Given the constrained domain, general purpose segmentation networks are overly computationally intensive. In this work, we train a generative segmentation network for semantic segmentation of oranges. The proposed method exploits the domain constraints to train a machine learning segmentation solution from scratch with only 110 labeled training images.