1. Grape counting in RGB videos – comparing two instance segmentation models

The number of grape berries per bunch between pea-size and bunch closure stages provides useful information to the farmer in planning and decision-making since it is an early indicator of the final yield to be harvested. The aim of this study is to count the number of grapes per cluster by comparing two different instance segmentation models (YOLACT and Spatial Embeddings) trained on RGB videos acquired with a UAV. YOLACT tends to undercount the number of grapes, with count estimations ranging from 0% to overcounts of 148%. Nevertheless, the lowest estimation achieved by Spatial Embeddings is 30% and the highest is 116%. In general, Spatial Embeddings segments and detects berries more accurately than YOLACT.