The Effects of Synthetic Data and Label Distribution on Canola Branch Counting with ResNet-18

Summary

This post summarizes an arXiv paper (2607.09630) by Amirsalar Darvishpour, Mikolaj Cieslak, and Adam Runions on computer vision for plant phenotyping. Collecting annotated plant images for automated phenotyping is slow and costly, so the authors systematically quantify two factors—synthetic-to-real data ratio and label distribution—when training ResNet-18 for canola branch counting. A 1:7 synthetic-to-real ratio yielded the best performance, reducing mean absolute difference by 7.6% versus real-only training. A uniform synthetic label distribution was strongly suboptimal; interpolating 90% toward the real distribution achieved an absolute difference of 0.927, while Gaussian smoothing of the real label distribution gave the best overall result (0.912, a 14.7% improvement over real-only training). A simple practical alternative is a minimum of 10 synthetic images per label, providing moderate gains.