Object Detection for Pepper Harvesting: A Comparative Evaluation of Backbone-Adapted, Contemporary, and Open-Vocabulary Models

Reliable object detection is essential for agricultural perception under illumination changes, occlusion, and complex backgrounds. This study investigates three complementary object-detection paradigms for multiclass pepper harvesting using 1004 RGB images annotated for human, pepper, pepper group, box, and pushcart. First, Faster R-CNN, SSD, and YOLOv7 were evaluated with alternative CNN backbones to examine the effect of feature extraction and training configuration. Second, the contemporary YOLO12n, YOLO26n, and RT-DETR-L detectors were evaluated under domain-specific closed-set training. Finally, YOLOE-26n was assessed under a zero-shot open-vocabulary setting without fine-tuning on the pepper-harvesting dataset. Among the backbone-adapted models, YOLOv7–GoogLeNet achieved 99.6% test [email protected]. Across five independent training seeds, YOLO12n achieved the highest mean performance among the contemporary detectors with 86.01±0.53% [email protected] and 71.00±0.62% [email protected]:0.95, while YOLO26n provided the most favorable efficiency profile at 47.06 FPS with 2.38 million parameters. Despite receiving no pepper-domain training, YOLOE-26n achieved 41.89% [email protected] and reached 92.43% [email protected] for human, demonstrating that pretrained visual–semantic representations can transfer effectively to selected categories in field-acquired agricultural scenes, although performance remained strongly class-dependent. Quantitative and qualitative analyses further highlighted differences in object recovery, localization, computational efficiency, and failure modes among the evaluated paradigms. Overall, the study provides a common experimental reference for understanding the trade-offs among backbone adaptation, contemporary closed-set detection, and zero-shot semantic recognition in pepper-harvesting perception. Because the three paradigms followed different training and benchmarking protocols, their results are interpreted within their respective experimental settings rather than as a strict cross-paradigm ranking.