Figure 4. Illustration of self-supervised learning by relative position prediction task.

(left): An image is divided into nine patches where the central patch (the one without number) represents the anchor patch and the remaining eight patches (delineated in dashed yellow lines) represent the query patches. (right): a training example that consists of an anchor patch and query patch is passed to a late-fusion convolutional model which share weights between the two branches to predict the position of the query patch with respect to anchor patch. Image credit: Gabriele Brancati.