Answers are grounded in this paper and cite their sources.
Instance segmentation is a key technology for precision agriculture tasks such as per-plant growth monitoring and harvest-timing prediction. However, in real smart-farm environments where crops are densely cultivated, instance segmentation poses two practical challenges. First, because pixel-level annotation is highly costly, only a small number of labeled images are typically available. Second, severe leaf overlap between adjacent plants causes boundary ambiguity that degrades segmentation accuracy. To improve segmentation performance, prior studies have explored incorporating structural priors of the target object into the learning process. However, existing methods require redesigned activation functions or separate optimization modules, making them hard to integrate with foundation models and largely limited to single-object semantic segmentation. To address these limitations, we propose ‘radial loss,’ an auxiliary loss function that explicitly captures the radial convex structure of rosette-shaped crops by supervising predicted foreground probabilities along rays cast from each instance's center of mass toward its boundary. We further introduce ‘dynamic search,’ a threshold-triggered sampling mechanism that focuses supervision on boundary regions once the validation mAP exceeds a predefined threshold. Experiments on three lettuce datasets show that, under an extreme 10-shot setting, the proposed method improves mean mAP by 7.7% points over the baseline, demonstrating its effectiveness in label-scarce, densely cultivated environments.
No takes yet. Share an insight, caveat, or question.
Lee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: