💻 computer science

The Effects of Synthetic Data and Label Distribution on Canola Branch Counting

This study demonstrates that training ResNet-18 models for canola branch counting with synthetic data generated by a calibrated L-system model significantly improves performance when the synthetic-to-real image ratio is optimized to 1:7 and the synthetic label distribution is smoothed to match the real data, achieving a 14.7% reduction in mean absolute difference compared to using real data alone.

Original authors: Amirsalar Darvishpour, Mikolaj Cieslak, Adam Runions

Published 2026-07-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Amirsalar Darvishpour, Mikolaj Cieslak, Adam Runions

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're trying to teach a robot to count the branches on a canola plant. In the real world, taking thousands of photos of plants and manually counting every single branch is a slow, boring, and expensive chore. It's like trying to learn a new language by only reading one very thick book.

To speed things up, scientists built a digital "plant factory" using a special code called an L-system. This factory can spin out unlimited pictures of fake canola plants, and because the computer built them, it knows the exact branch count for every single image. It's like having a magic textbook where every answer is already written in the back.

But here's the catch: just because the computer knows the answers doesn't mean the robot learns them easily. The fake plants look a little too perfect compared to the messy, real ones. The big question was: How do we mix these fake photos with real ones so the robot gets really good at counting?

The researchers ran a series of experiments to find the perfect recipe, and they discovered three golden rules.

1. The "Too Much of a Good Thing" Ratio

First, they tested how many fake photos to mix in with the real ones. Think of it like seasoning a soup. If you add a pinch of salt (fake data) to a pot of soup (real data), it tastes better. But if you dump in the whole shaker, it becomes inedible.

They found that adding fake photos helps a lot, but only if you don't go overboard.

The Sweet Spot: When looking strictly at the mix of fake-to-real photos, the best ratio was 1 fake photo for every 7 real photos (a 1:7 ratio). This cut the robot's counting errors by 7.6% compared to using only real photos.
The Safe Zone: You don't need to be a math genius to get it right. Any mix between 1:5 and 1:22 worked well.
The Danger Zone: If you go past 1:22 (meaning too many fake photos), the robot starts to get confused by the "fake" look of the images, and its performance actually gets worse than if you had just used real photos alone.

2. The "Fake Crowd" Problem

Next, they looked at which fake plants to generate. In their digital factory, the plants grew in a way that created a "uniform" distribution. This means the factory made an equal number of plants with 1 branch, 10 branches, 20 branches, and 40 branches.

In the real world, however, most plants have a moderate number of branches, and very few have extreme numbers (like 1 or 40).

The Mistake: Using a "uniform" mix where every branch count is equally common was a disaster. It made the robot's errors jump to 1.70 (a very bad score). It's like trying to learn about the weather by assuming it rains, snows, and is sunny with equal frequency every day.
The Fix: They tried to make the fake plants look more like the real crowd. They found that if they shifted the fake distribution to be 90% similar to the real distribution, the errors dropped to 0.927.
The Magic Trick (Gaussian Smoothing): The absolute best result of the entire study came from a technique called "Gaussian smoothing." Imagine taking the real data and gently blurring the edges so that rare branch counts get a little extra attention, but without making the whole thing look fake. This simple tweak brought the error down to 0.912, a 14.7% improvement over using real photos alone. This was the true winner, beating even the best ratio mix.

3. The "Minimum Guarantee" Trap

Finally, they tested a simpler idea: "What if we just promise to have at least X fake photos for every single branch count, even the rare ones?"

The Modest Approach: If they guaranteed at least 10 fake photos for every branch count, the robot did okay (errors dropped to 0.993). It wasn't as good as the "smoothing" trick, but it was a safe, simple backup plan.
The Over-Correction: If they tried to guarantee 100 fake photos for every count, the robot got confused again, and errors rose to 1.109. By forcing too many rare examples, they accidentally made the fake data look too uniform again, ruining the progress.

The Bottom Line

The researchers didn't just guess; they measured these results five times each to be sure. They found that while the exact mix of fake-to-real photos matters, it's forgiving—you don't need to be perfect. However, how you arrange the fake data matters a huge amount.

The best strategy isn't to just dump random fake plants into the training set. Instead, you should take the real data, gently smooth it out to give rare branch counts a little boost (using that Gaussian trick), and mix in about 1 fake photo for every 7 real ones. This approach helps the robot bridge the gap between the digital world and the real world, making it a much better counter without needing to take millions of expensive photos.

While this study was done specifically on canola plants, the scientists suggest these rules might help with other crops too, though they admit they still need to figure out exactly which parts of the digital plant model are the most important for making this work.

Technical Summary: The Effects of Synthetic Data and Label Distribution on Canola Branch Counting

Problem Statement
Automated plant phenotyping via supervised deep learning typically requires large, annotated datasets, which are costly and time-consuming to produce. While procedural L-system plant models can generate unlimited synthetic images with exact organ-count labels, previous research indicates that simply adding synthetic data does not guarantee performance improvements. The efficacy of synthetic data incorporation depends critically on two factors: the ratio of synthetic to real images and the label distribution of the synthetic dataset. This paper addresses the need to systematically quantify how these two factors influence model performance in the specific context of counting canola branches.

Methodology
The authors utilized a calibrated L-system plant model to generate synthetic canola images, ensuring the synthetic branch counts qualitatively matched real-world observations. The study employed a ResNet-18 model (pre-trained on ImageNet) fine-tuned for branch count regression. The evaluation was based on a dataset of 788 real canola images with ground-truth branch counts (ranging from 1 to 40).

The experimental design isolated two primary variables across four distinct experiments:

Ratio Sweep: The authors varied the synthetic-to-real image ratio from 1:0 to 1:29 (specifically testing 30 ratios) while keeping the real dataset fixed.
Distribution Interpolation: The synthetic label distribution was shifted in ten equal steps from a uniform distribution (Step 0) toward the real data distribution (Step 10). This was tested with 18k and 36k synthetic images paired with 400 real training images.
Gaussian Smoothing: A Gaussian kernel was convolved with the real label distribution histogram to smooth branch-count frequencies across neighboring values. Ten values of $\sigma$ (from 0.0 to 3.6) were tested.
Minimum Samples per Label: Hard-floor constraints were applied to ensure a minimum number of synthetic images per label value (10 and 100 images per label), aiming to up-sample rare high-count labels without requiring full knowledge of the target distribution.

All experiments were repeated five times, with performance evaluated on a held-out test set using the Mean Absolute Difference (abs. diff.) as the primary metric.

Key Results

Synthetic-to-Real Ratio: The study identified a broad "beneficial range" for the synthetic-to-real ratio between 1:5 and 1:22. Within this range, performance generally improved over real-only training. The optimal ratio was found to be 1:7, which reduced the mean absolute difference by 7.6% (from 1.069 to 0.988) compared to real-only training. Ratios exceeding 1:22 resulted in performance degradation, where the excess synthetic data amplified domain-gap noise.
Label Distribution (Interpolation): A uniform synthetic distribution was found to be strongly suboptimal (abs. diff. $\approx$ 1.70). Interpolating the synthetic distribution 90% toward the real distribution (Step 9) yielded an abs. diff. of 0.927, outperforming even the real distribution itself (0.942). This suggests that slightly over-representing rare labels and smoothing frequency extremes improves generalization.
Label Distribution (Gaussian Smoothing): Applying Gaussian smoothing to the real label distribution yielded the best overall result across all experiments. At $\sigma = 3.6$ , the model achieved an abs. diff. of 0.912, representing a 14.7% improvement over real-only training. The performance curve flattened for $\sigma$ values between 2.0 and 4.0.
Minimum Samples: Enforcing a minimum of 10 synthetic images per label (min-10) provided modest gains (abs. diff. 0.993) over real-only training. However, enforcing a minimum of 100 (min-100) hurt performance (abs. diff. 1.109), as aggressive up-sampling of rare labels shifted the distribution toward uniformity in the tail, reintroducing the issues associated with uniform distributions.

Significance and Contributions
The paper provides a systematic characterization of synthetic data design choices for plant phenotyping. Its primary contributions are threefold:

Ratio Forgiveness: It establishes that the synthetic-to-real ratio is broadly forgiving; practitioners do not need to tune the ratio precisely, as any value between 1:5 and 1:22 outperforms real-only training.
Distribution Optimization: It demonstrates that label distribution has the largest impact on performance. Specifically, avoiding uniform distributions and applying Gaussian smoothing to the real label distribution (or interpolating 90% toward the real distribution) significantly closes the synthetic-to-real domain gap.
Practical Alternatives: It offers a simpler alternative for scenarios where distribution resampling is not viable: setting a modest minimum floor of 10 images per label. However, it cautions that higher floors (e.g., 100) can be detrimental.

The authors conclude that while model construction and calibration require substantial effort, augmenting real images with synthetic data—when paired with appropriate distribution matching—makes combined datasets superior to real-only baselines. The findings suggest that simple adjustments to the observed label distribution can substantially improve model performance, though the authors note that further investigation is needed to determine if these results generalize to other crops and which specific components of the L-system model are most critical for bridging the domain gap.

1. The "Too Much of a Good Thing" Ratio

2. The "Fake Crowd" Problem

3. The "Minimum Guarantee" Trap

The Bottom Line

Technical Summary: The Effects of Synthetic Data and Label Distribution on Canola Branch Counting

More like this