Split before you celebrate accuracy

The move: split by the thing that creates look-alike images. A train-validation-test split is not bookkeeping. It is your estimate of how the model handles images it has not effectively seen before. In computer vision, random image splits can be misleading because many images are correlated. A video produces many frames that differ only slightly. A store camera produces repeated angles. A document scanner produces the same template. Medical images from the same patient share anatomy and acquisition conditions. If correlated images cross splits, the model can perform well by memorizing the scene, background, camera, or template. The validation score…

Sign up free — one personalized lesson every day, matched to your role and goals.

Already have an account? Sign in