SynthAug and SNMS: Enhancing Flower Detection via Data Augmentation and Improved NMS

• SynthAug: memory-efficient data augmentation for resource-limited training. • Second Non-Maximum Suppression: reduces false positives via geometric clustering. • Increased pollination coverage and efficiency in real-world robot experiments. Effective deployment of autonomous agricultural robots for tasks such as pollination and yield estimation requires robust visual perception. However, the critical task of flower detection is fundamentally challenged at two stages: during training, by data scarcity resulting from high-cost annotation, and during inference, by the high volume of redundant detections that hinders model precision. These challenges lead to unreliable robotic decisions, unnecessary actions, and reduced coverage, thereby lowering operational efficiency and limiting agricultural output in field environments. Conventional data augmentation methods are often memory-intensive, while standard Non-Maximum Suppression (NMS) struggles to filter redundant predictions without sacrificing recall. To address these dual challenges, we propose two targeted solutions: SynthAug and Second Non-Maximum Suppression (SNMS). SynthAug is a memory-efficient augmentation technique that enriches training data by pasting segmented foreground instances onto individual images. SNMS is a post-processing algorithm that adaptively removes redundant detections by grouping them based on spatial proximity and iteratively evaluating their confidence scores. Experiments on two custom flower datasets demonstrate that SynthAug improves AP 50 by an average of 1.5% over robust baselines while reducing memory consumption by 79% and increasing processing speed by 34%. SNMS reduces false positives by 53% with negligible impact on recall or inference speed. Crucially, real-world robotic pollination experiments validate our approach: SynthAug and SNMS collectively increase pollination coverage by 6.0% and reduce the average time per pollination by 28%. These results demonstrate that our methods effectively bridge the gap between model training and practical deployment. These methods enhance model performance given limited training data while maintaining a precise balance between precision and recall, thereby reducing deployment costs and improving both coverage and operational efficiency for agricultural robots.