Paintings
800 crops from paintings, each showing one of ten subjects — birds, boats, chairs, cows, dogs, horses, sheep, trains and dining tables.
Classes
- bird 63
- boat 201
- chair 79
- chair diningtable 39
- cow 47
- diningtable 76
- dog 80
- horse 123
- sheep 63
- train 29
The mount around each thumbnail is its class. Nothing else is encoded — the picture is the data, and the colour is only there so you can see whether the projection grouped the classes or not.
Denser, and harder
Section titled “Denser, and harder”Half again as many images as the flowers, over a projection with much less to hold on to. Photographs of flowers vary mostly in colour; paintings vary in style, palette, period and brushwork before they vary in subject, so a projection computed from image content groups how a thing was painted at least as strongly as what was painted.
The result is a plot where the classes interleave, and where reading the arrangement means reading the pictures. That is a case the glyph showcases cannot make: no encoding of ten categorical labels would tell you that two neighbouring crops share a muted ochre palette rather than a subject.
What the grid costs here
Section titled “What the grid costs here”More than usual. The projection is dense in the middle and sparse at the edges, so a lattice pulls the centre apart hard — watch the displacement readout as you switch layouts. The Lens is the other answer to exactly this: leave the projection alone and grid only what is under the pointer.
References
Section titled “References”The State of the Art: Object Retrieval in Paintings using Discriminative Regions British Machine Vision Conference 2014. doi:10.5244/C.28.38
The paintings these crops come from, labelled with the subject each depicts.
Hagrid: using Hilbert and Gosper curves to gridify scatterplots Journal of Visualization 25(6), 1291–1307, 2022. doi:10.1007/s12650-022-00854-7
Where this example first appeared.