One of our client, a Fortune 500 company has delegated us the task to train its algorithm of Image Quality Scoring, for understanding the quality of the diffusion of their product images circulating on the Internet, especially on the product pages of marketplaces or ressellers.
Stackadoc helped them go from idea to production and control the quality of the results of our algorithm
Automatic image evaluation of online marketing assets is important for various reasons
We created a fully-customized Data Science platform from data collection to business result activation going all the way through models training and deployment. The core of the AI used transfer learning from existing image analysis algorithm that we fine-tuned in order to improve the degree of the quality detection, including physical criteria as well as human-opinion-based esthetical criteria.
We packaged the whole process into an API that could also be consumed by our client as well as by the interface that we resote
As of today, we control daily 10,000s of packshots that our robots harvest from marketplaces accross the globe. We can process 100s of images per second and thus ensure that the quality of the content that is delivered is of optimal quality. Audits and A/B test realised by our client tend to show the ROI of the tools is largely higher than expected.