The 656 patches of the test fold: the models never saw them in training. The strip under a patch shows its true label on the left and what the three models say together on the right; stripes mean they are below the confidence threshold.
A picture with its test patches, outlined in the class the three models give each one (dashed when below the threshold). Select a patch to see it; switch patches off to see how the call for the whole picture changes.
The picture cut into a grid of square tiles, each classified by the three models together: a tile takes the colour of its class, and stripes when the models are below the threshold. Select a tile, or move from tile to tile with the arrow keys, to see it and what each model says. Where a tile is one of the labelled patches, its true label is shown.
Draw a box anywhere on the picture to classify that crop. A box that is not square, or not 512 pixels, is stretched to the models' input: they were trained on square patches of 512.