Abstract

Manual assessment of soft fruits is both laborious and prone to human error. We present methods to compute width, height, cross-section length, volume and mass using computer vision cameras from a robotic platform. Estimation of phenotypic traits from a camera system on a mobile robot is a non-destructive/invasive approach to gathering quantitative fruit data which is critical for breeding programmes, in-field quality assessment, maturity estimation and yield forecasting. Our presented methods can process 324–1770 berries per second on consumer-grade hardware and achieve low error rates of 3.00 cm3 and 2.34 g for volume and mass estimates. Our methods require object masks from 2D images, a typical output of segmentation architectures such as Mask R-CNN, and depth data for computing scale.