Lychee Instance Segmentation at Different Growth Stages Using YOLOv8-seg Model

This paper presents an approach to lychee instance segmentation through different development stages using the YOLOv8-seg model. A new dataset comprising 1006 lychee images from various growth phases (flowering, young fruit, green fruit, and ripe fruit) was introduced and manually labeled. We tested 5 versions of YOLOv8-seg on this dataset. The findings reveal that YOLOv8x-seg, with 71.8M parameters, achieves the highest average mAP50 accuracy of $84.1 \%$, while YOLOv8n-seg, with only 3.4M parameters, attains an mAP50 of $83.7 \%$. These results suggest that YOLOv8n-seg suits real-world applications on less powerful hardware.