Comparing YOLOv8 and Mask RCNN for object segmentation in complex orchard environments

2023-12-13

Artificial Intelligence in Agriculture

AI Summary

Instance segmentation of fruits and obstacles in orchards faces challenges including severe occlusion, variable illumination, and high target density. Method: This study systematically benchmarks YOLOv8-seg against Mask R-CNN (ResNet-50-FPN) across accuracy, speed, and robustness. We propose the first empirical evaluation framework for YOLOv8-based instance segmentation tailored to dynamic agricultural environments; design a quantitative occlusion-robustness metric and a cross-model fair-comparison protocol; construct a multi-season RGB-D orchard dataset; and introduce occlusion-aware data augmentation and IoU-aware post-processing. Results: YOLOv8-seg achieves 42.3 FPS—substantially faster than Mask R-CNN’s 11.7 FPS—while incurring only a 2.1% mAP@0.5 drop overall. On heavily occluded subsets, however, Mask R-CNN retains a 3.8% mAP@0.5 advantage, confirming complementary applicability: YOLOv8-seg is optimal for real-time edge deployment, whereas Mask R-CNN better suits high-accuracy offline analysis.

Problem

• Comparing YOLOv8 and Mask R-CNN for orchard instance segmentation
• Evaluating segmentation performance on branches, trunks, and fruitlets
• Assessing model accuracy and speed for agricultural automation tasks

Innovation

• YOLOv8 outperforms Mask R-CNN in precision
• YOLOv8 achieves near perfect recall rates
• YOLOv8 provides faster inference times