Answers are grounded in this paper and cite their sources.
Accurate fruit detection in cluttered orchard environments remains challenging due to visually similar background elements such as foliage and specular highlights, which lead to high false positive rates in RGB-based detectors. In addition, conventional RGB-D fusion methods often degrade under real-world depth noise and domain shifts. To address these limitations, this study introduces GeoSphere-DETR, a geometry-aware extension of RT-DETR that leverages robust depth-derived cues for reliable fruit detection. A real-world RGB-D citrus dataset comprising 2200 images and 102,719 annotated instances was collected across multiple orchards and acquisition days, with a strict cross-field and cross-day evaluation protocol to assess generalization. An efficient annotation pipeline combining a foundation model (Grounding DINO) with human verification was developed to reduce labeling effort while maintaining annotation quality. GeoSphere-DETR integrates two key components. First, a neck-level geometry-gated fusion module incorporates curvature-based geometric cues (bumpiness and Laplacian-based convexity) into multi-scale visual features via learnable spatial gates, enabling selective use of reliable depth information. Second, a query-level geometry-aware recalibration head re-scores detection confidence based on geometry-informed confidence adjustment, suppressing geometrically inconsistent false positives. Experimental results on the held-out test set show that GeoSphere-DETR achieves a mAP50 of 0.863 and 3.29 FPPI, outperforming RGB-only RT-DETR (0.786 mAP50, 7.24 FPPI) and YOLOv11m (0.715 mAP50, 8.735 FPPI), corresponding to a 9.8% improvement in mAP50 and a 54.5% reduction in FPPI over RT-DETR. Ablation studies confirm the effectiveness of curvature-based cues and gated multimodal fusion. Overall, this work introduces a dual-level geometry-aware detection framework that combines feature-level reliability modeling with object-level shape consistency, providing a principled approach for reducing false positives in RGB-D agricultural perception systems and building a perception foundation for robotic harvesting, yield estimation, and precision orchard management. The dataset and code are publicly available at citrus-fruit-rgbd ( https://huggingface.co/datasets/ariel0818/citrus-fruit-rgbd ) and GeoSphere-DETR ( https://github.com/Ariel0818/Geosphere-DETR ), respectively.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: