Esoft’s data science department has compared Esoft’s AI model to other world-class AI projects. The result? Esoft’s model is among the most precise in the world.

Most likely, few people outside of the narrow sphere of AI academics have heard of the MIT-Adobe FiveKdataset. But for AI and computer vision scholars around the world, the database is a very valuable tool used to verify the accuracy and precision of AI models.

That is also the case for the tech masterminds in Esoft’s data science department, who have used the database to benchmark Esoft’s AI model up against other similar projects.

The results have been extremely uplifting.

We have found that our model is actually the most precise when comparing it to other top 10 projects in the database,” says Juan Francisco Marin Vega, an industrial Ph.D. candidate at Esoft.

5,000 high resolution images

The MIT-Adobe FiveK dataset, initiated by elite tech university Massachusetts Institute of Technology (MIT), is an open-source database consisting of 5,000 image pairs in 4K resolution – one of the largest publicly available collections of high-resolution image data in the world.

The database comprises of both input and output data – i.e. a combination of raw image files (input images) and JPG output images, which have been manually edited by professional photo editors.

“As a researcher or developer, you can thus verify the accuracy and precision of your model by selecting a certain amount of input images, training your model on them and then comparing the results to the output images in the database,” Marin Vega explains.

He adds that for a company like Esoft, the FiveK dataset is a very useful tool. While Esoft has access to millions of its own data, these data are restricted by copyright issues and cannot be shared – thus making it difficult to validate the results.

The importance of zero errors

The model training sessions have been used to gradually improve Esoft’s model. This model serves as part of the underlying framework for the AI solutions currently being implemented in Esoft’s production.

The theoretic results from the FiveK dataset training can of course not be transferred 1:1 over to the commercial side of the business. But for Esoft’s Head of Research & Development, Kasper Grud Skat Madsen, these kinds of scientific approaches are nonetheless of great importance for commercial development.

“In order to use these AI tools in our everyday deliveries, we need to make sure that there are as few errors as possible,” says Grud Skat Madsen. “Therefore, it is very reassuring to see that our models are best-in-class when compared to similar projects out there.”