Computer Vision News - May 2023

33 Marcel and Miguel is already annotated . To achieve this, the team has developed a similarity estimator consisting of an autoencoderwith a specific loss function and arrangement. Miguel believes their method is more objective than relying solely on human assessment. It considers various factors essential for determining dataset similarity – not only the modality of the data but also the style. “ The style means the aspect, color, type of annotation and types of organs or cells represented in the image, and it considers all these things as a whole to match the datasets, ” he explains. “ Humans can see by sight if a database is similar, but they don’t know if it’s correlated with the pre- training results. We’ve demonstrated that our similarity estimator corresponds to the pre-training results in our semantic segmentation task. ” The new method is designed to work with any 2D image dataset , making it applicable to a wide range of medical and non- why we’re working on these data-related problems and can see developments toward a more data-centric AI . We believe focusing more on the data and less on the architectures is important. ” The idea behind the approach is to automate the selection of suitable databases for a given problem by identifying themost similar database that Miguel Molina-Moreno Marcel Schilling