Beyond green: Domain-aware self-validated instance counting for loropetalum and non-green ornamental species using annotation-free robustness metrics

Authors: MANJUNATHA, H - The University Of Texas At Dallas; SUNDARAVADIVEL, P - University Of Texas; BORAH, S - University Of Texas; TAMIL, L - The University Of Texas At Dallas; KNIGHT, P - Mississippi State University; Torbert Iii, Henry; Kumpatla, Siva Prasad

Submitted to: IEEE International Conference on Computer Vision and Pattern Recognition
Publication Type: Proceedings
Publication Acceptance Date: 3/20/2026
Publication Date: 6/3/2026
Citation: Manjunatha, H., Sundaravadivel, P., Borah, S.S., Tamil, L., Knight, P.R., Torbert III, H.A., Kumpatla, S. 2026. Beyond green: Domain-aware self-validated instance counting for loropetalum and non-green ornamental species using annotation-free robustness metrics [abstract]. IEEE International Conference on Computer Vision and Pattern Recognition.

Interpretive Summary: The use of advanced technologies in agriculture has brought about a major shift toward precision agriculture, a data-driven approach focused on improving crop yield, reducing resource use, and promoting sustainability. Unmanned Aerial Vehicles (UAVs) have become especially valuable for large-scale data collection at field level. However, counting and segmenting densely packed objects in challenging field conditions remains an open problem in computer vision, especially for UAV-based agricultural monitoring. Traditional accuracy-centric metrics often fail to reveal a model’s reliability when confronted with dense occlusion, non-standard coloration, and variable lighting. This manuscript describes a new evaluation framework comprised of four complementary metrics: Radial Counting Stability (RCS), Cross-Scale Consistency (CSC), Semantic-Visual Stability (SVS), and Adaptive Repeatability Index (ARI), each quantifying a distinct form of model consistency. This new technology was tested on 800 purple-foliaged ornamental samples of orthomosaics of Loropetalum chinense canopies. The overall framework enables practical, annotation-free assessment of model readiness for deployment in agricultural and nursery applications.

Technical Abstract: Counting and segmenting densely packed plants in real field conditions is still a challenge for UAV-based agricultural vision. Traditional accuracy metrics often fail to reflect true reliability under occlusion, color variation, and lighting changes. This work introduces an annotation-free evaluation framework that measures robustness without ground-truth labels using four complementary metrics: Radial Counting Stability (RCS), Cross-Scale Consistency (CSC), Semantic-Visual Stability (SVS), and Adaptive Repeatability Index (ARI). Each captures a different aspect of model consistency directly from predictions, allowing fast and interpretable validation. Tested on 469 high-resolution UAV images of Loropetalum chinense (over 2,600 plants per frame), results showed that even models with >91% accuracy can exhibit poor robustness (CSC < 0.80). The proposed metrics strongly correlate with deployment success (' = 0.87, p < 0.001) and reduce validation time by about 15× compared to manual annotation. Cross-dataset results on Mangifera indica (RCS = 0.92 ± 0.04) and Euphorbia esula (' = 0.71, p < 0.01) confirm generalizability, while tests on 800 purple-leaved ornamentals demonstrate adaptability to pigment variation, validating their effectiveness for real-world agricultural applications. The resources along with the demo are available at: https://github.com/harshitha-8/Beyond-Green-Loropetalum-and-Non-Green-Ornamental-Species-Annotation-Free-Robustness-Metrics.