Greenwood Engineering, based in Denmark, develops advanced measurement systems used across road, rail, and airport infrastructure. Their equipment continuously records high-resolution road-surface images, generating extremely large datasets that capture the texture, condition, and marking patterns of road networks at scale.

As the volume of captured data grew into the millions, Greenwood began exploring machine learning approaches to classify surface types, detect patterns, and support automated quality assessment. 

While collecting data wasn’t an issue, making it useful was the real challenge.

Greenwood Engineering trains models for detecting and measuring road surface defects, such as cracks and potholes, as well as measuring lane marking quality. But labeling millions of images was infeasible, and manually selecting which samples to label was slow, repetitive, and prone to redundancy:

• Many images captured nearly identical road segments.
• Important edge cases were buried in the dataset.
• Labeling at scale was costly, and manual filtering didn’t scale.

Labeling this entire corpus was not feasible, and manual sampling lacked consistency.

To efficiently curate their dataset, Greenwood used LightlyTrain to train their own DINOv2 model on unlabeled road surface images. The resulting model captures different road surface conditions much better than an off-the-shelf model. This made data curation the most effective lever for improving model performance, and LightlyTrain enabled that shift. 

Using LightlyTrain and the custom DINOv2 model, the team generated embeddings for their entire dataset. These embeddings gave them a scalable way to explore the data, run similarity search, remove redundancy across millions of road-surface images, and extract valuable samples for labeling.

Why Curation Was Essential

With the current dataset, the team quickly reached the point where finding relevant samples for labelling required manually inspecting hundreds or thousands of samples.

What they needed instead was a better understanding of which images were actually informative. LightlyTrain helped provide that structure:

• Training improvements from additional labeled data were modest
• Adding labels without addressing redundancy led to diminishing returns
• The dataset needed to be organized before annotation could have impact

LightlyTrain helped Greenwood organize their large road-surface dataset into a meaningful embedding space. This gave the team a much clearer understanding of where redundancy existed, which textures and markings were visually similar, and which samples were diverse enough to prioritize for labeling.

With this visibility, annotation could focus on examples that were most likely to improve downstream models.

• Identify clusters of similar road textures and markings
• Run k-nearest-neighbor (kNN) queries to find visually related samples
• Detect redundant images across long road stretches
• Build a more balanced and representative labeled subset