Reproducible Benchmarking of Tomato Detection in Greenhouse: Comparing Attention-Augmented and Baseline Detectors

Accurate tomato detection in greenhouse imagery is essential for robotic harvesting, yield estimation, and crop monitoring, yet visual clutter, fruit overlap, partial occlusion, and variable illumination remain challenging for object detectors. Although attention modules are frequently used in agricultural vision studies to improve feature discrimination, their practical contribution is often reported without controlled comparison against strong baseline detectors. This study presents a reproducible and deployment-aware benchmark for single-class greenhouse tomato detection using 895 images with 4930 annotated tomato instances in PASCAL VOC format. The first experimental block used a fixed 70/20/10 split to compare Faster R-CNN, four attention-augmented Faster R-CNN variants, Cascade R-CNN with ResNet101-DCN-FPN, and YOLOv11s attention variants. A second extended protocol converted the annotations to YOLO format and evaluated YOLO-family detectors and RT-DETR-l under a stratified 70/15/15 split, including ablation, robustness, seed-stability, and deployment analyses. The annotation audit confirmed valid bounding boxes, no empty images, and a high proportion of small tomato instances. In the first block, attention integration did not consistently improve detection performance, whereas Cascade R-CNN achieved the highest accuracy with 92.80% ${\mathrm{mAP}}_{0.5}$ and 90.80% $F_1$-score. In the extended protocol, RT-DETR-l obtained the highest test accuracy with 91.49% ${\mathrm{mAP}}_{0.5}$ and 58.59% ${\mathrm{mAP}}_{0.5:0.95}$, while Final-YOLO11s achieved comparable performance with lower latency, reaching 91.42% ${\mathrm{mAP}}_{0.5}$, 58.37% ${\mathrm{mAP}}_{0.5:0.95}$, and 86.19% $F_1$-score. Across three seeds, Final-YOLO11s obtained a stable mean ${\mathrm{mAP}}_{0.5}$ of 90.84%. Robustness analysis showed that motion blur and Gaussian noise caused the largest degradation, whereas compact YOLO models exported reliably to ONNX and TensorRT. Overall, the results indicate that localization quality, robustness, latency, model size, stability, and export capability should be considered together, and that adding attention modules by default is less reliable than evidence-driven detector selection.