Burberry builds object detection and classification models that predict which marketing images will perform. Labelbox produces the visual signal, inside Databricks, turning two-month analyses into two-hour self-service.

Originally published from the Databricks website. Source: https://www.databricks.com/customers/burberry

The challenge

Burberry, the British luxury brand, runs global marketing campaigns on thousands of images from many sources. To use those assets well, it has to classify them precisely — and it trains object detection and classification models to do it. Producing that signal by hand wasn't feasible, and the models had to identify highly specific products.

We work with high-volume unstructured data sets to train our object detection and classification models,” explained Harry Collings, Data Science Manager at Burberry. “It’s simply not feasible to identify and label all those images manually. And because our image recognition models must identify highly specific products, we eventually realized we needed a dedicated solution that could work precisely while taking the manual effort out of the process.

Burberry first tried an open-source annotation tool. It worked, but with serious drawbacks.

Our open source tool lacked a proper interface that could be read from our data sources,” Collings recalled. “All the images had to be stored locally on a data scientist’s machine, which was not ideal for a company of our size. To label, say, 1,000 images would take someone two hours — and they could only save their progress in a JSON file. We knew it was time to look for a more robust commercial solution.

The approach

Burberry implemented Labelbox inside its Databricks Lakehouse Platform, treating images like any other dataset. It connected Labelbox to its virtual private cloud and S3, imported master assets via API, and was up and running within a month. Labelbox produces the visual signal that trains Burberry's models, and its Model product finds and fixes the label errors that most affect results.

The lightbulb moment was when we saw how easily we could import our images to Labelbox from our Amazon S3 bucket via API,” reported Collings.

The outcome

Generating insights on images used to take two months of hands-on data science work; on Labelbox it's a two-hour self-service process. Marketing users upload candidate images into Burberry's in-house Content Ranking Engine (CRANE) and score them against pretrained models that predict the engagement images will get in upcoming campaigns by channel and region — a ranking that informs creative briefs and decisions.

We’re not only helping our stakeholders choose the best images to use in campaigns but also tracking the revenues generated by the emails they send,” said Collings. “This information influences each batch of imagery we produce and leads to continuous improvement in our marketing programs.

I would have had to ask for a 10-person team just to build a product that would let us view all our imagery, let alone generate insights on it,” Collings remarked. “In our day-to-day workflow, Labelbox saves us countless hours while delivering a tremendous amount of value.

With Databricks Lakehouse Platform and Labelbox, we’re able to realize our full ambition of creatively led, data-driven content creation. Labelbox continually gives our team insights on how to optimize our campaigns. And Databricks helps us maintain an efficient, sector-leading analytics capability in the cloud.

– Ben Halsall, Analytics Technology Director, Burberry

Where this goes

Every campaign produces new signal, and every model gets better for the next one. That's a marketing learning loop — creative judgment, captured as structured signal, compounding into better decisions.