Cross-Generation Optimization of YOLOv26, YOLOv11, and YOLOv8 for Fine-Grained Small-Object Detection and Instance Segme
A new arXiv paper benchmarks Ultralytics YOLOv8, YOLOv11, and YOLOv26 models for detecting and segmenting small apple structures—fruitlets, calyx, and peduncle—in orchard settings. The work focuses on fine-grained instance segmentation under conditions of green-on-green camouflage, occlusion, and limited pixel detail for each target.
The evaluation addresses perception tasks directly relevant to agricultural robotics, where accurate identification of these structures supports automated operations such as thinning, harvesting, and crop monitoring. Results from the cross-generation comparison supply performance data on the same dataset and task definitions.
The study records model outputs for both detection and segmentation metrics without claiming superiority of any single architecture outside the tested orchard scenarios.