Effects of Synthetic Data Ratio and Label Distribution on Canola Branch Counting with ResNet-18 (arXiv 2607.09630)

Summary

This paper by Amirsalar Darvishpour, Mikolaj Cieslak, and Adam Runions (arXiv:2607.09630, posted 2026-07-10) systematically quantifies how two factors—the proportion of synthetic images and the label distribution—affect deep learning performance in plant phenotyping. Using canola branch counting as the benchmark task and a ResNet-18 backbone, the authors find that mixing synthetic and real images at a 1:7 ratio yields the best performance, reducing mean absolute difference by 7.6% compared to training on real data alone. Label distribution proved equally important: a uniform synthetic label distribution is strongly suboptimal, while interpolating 90% toward the real distribution achieves an absolute difference of 0.927. Gaussian smoothing of the real label distribution delivers the best overall result, with an absolute difference of 0.912, a 14.7% improvement over real-data-only training. As a practical alternative, using at least 10 synthetic images per label provides modest gains with a simple heuristic. The work offers actionable guidance for researchers facing the high cost of collecting annotated plant imagery.