Tomato Maturity Analysis: A Comparative Study of Detection and Instance Segmentation Using YOLOv8

The accurate visual analysis of fruit maturity in complex agricultural scenes remains a fundamental challenge due to gradual appearance changes, object overlap, and partial occlusion. This study addresses tomato maturity analysis, formally defined as instance-level binary classification and spatial localization under varying degrees of visual density. While bounding-box-based object detection is widely used, it often lacks precision in dense clusters. We present a controlled experimental comparison between object detection and instance segmentation using a common YOLOv8-medium (YOLOv8m) backbone to isolate the effect of spatial representation. Experimental results demonstrate that instance segmentation achieves superior localization accuracy and boundary consistency, reaching a mask-based mAP@0.5:0.95 of 0.817. These findings suggest that pixel-level supervision effectively reduces localization ambiguity, providing a robust foundation for automated agricultural monitoring.