For us, a big part of the challenge of making machine learning practical has been shrinking the cycle time between idea and production. We knew we were making progress when we could rapidly rebalance our training data and integrate feedback from our ML team into the process. The iterative cycle sped up dramatically, and we could move our models from idea to production in record time.
– Mark Lemmons, co-founder of Shotzr
In a world with nearly 10 billion screens, customer experience and attention are driven through imagery. Shotzr helps you identify your customer, and then identify with them through contextual imagery that matches your audience targeting.
Whether you’re doing Facebook and Instagram campaigns, social media marketing, SEM landing pages, or truly dynamic creative, Shotzr provides you with all the imagery you need for digital marketing.
Shotzr’s collection of curated, high-quality imagery is growing at a breakneck pace with access to nearly 100M images. In order to deliver the exact image marketers need for every marketing moment, Shotzr’s images need to be properly and accurately annotated.
As Shotzr’s contributors upload more images, and Shotzr’s partnerships with other image providers continue to grow, the in-house team could not keep up with labeling and metadata demands, making it difficult to scale. Shotzr annotated approximately 20,000 images over 90 days, but that wasn’t going to scale to the massive amounts of images on the platform.
To ensure Shotzr’s platform serves up the right images that marketers are looking for, Shotzr needed to create a solution identifying images that needed specific labels. These labels display only the shortlist of perfect options to customers, as opposed to scrolling through hundreds of images that match the search query. For example, some images are clearly from a specific location – like Aspen, Colorado. Labeling this location can help people trying to find specific photos of Aspen, not the trees, but the town and surrounding areas. But some images don’t have a distinguishable location – like a photograph of a bunch of hands or a typical stock photo.
By turning to machine learning, Shotzr wanted to speed up and automate the process of identifying which images needed to have specific location metadata. Proving value early on and actually delivering that value with ML was critical to Shotzr’s success with moving from pilot to production because of their initial limited resources for making the model work. To prove that value, Shotzr was relying on obtaining high-quality data to properly and accurately train their models.
To offer that level of granular labeling, Shotzr wanted a data partner that was up to the challenge of the volume of work, for a company their size, and with rational pricing. This is why they turned to us for help with the data that fed into their search relevance model.
Fast forward to only a few weeks later, and Shotzr is anticipating training 4x as many classifiers as they thought possible.
After their first job in the Appen platform, Shotzr identified over 17,000 images that did not require additional labeling. They anticipate over 61 million assets that they can remove from consideration for location data, freeing up their time to focus on images that can benefit from location data, as well as creating new models to automate the location labeling process.