A machine can analyse Wes Anderson's visual motifs better than you

Wes Anderson’s films are famous for their colour palettes, specific humour and captivating stories, motifs that create a cohesive and incredibly distinctive body of work. However, what exactly are these motifs and how do they manage to create films that are so specifically “Wes Anderson”? We, thanks to the work of Yannick Assogba, we can now pick apart his films, beyond the usual symmetrical compositions and surreal themes with the help of Inception V3, a “deep neural network” produced by Google.

Yannick Assogba is the software developer and designer working at the intersection of computation and design who, having been a fan of Anderson’s for years decided to utilise Inception V3 in order to unpack some of his favourite films: The Life Aquatic, The Royal Tenenbaums, Fantastic Mr. Fox and Moonrise Kingdom.

The process involved extracting one frame from each film every ten seconds, producing a sample of 2,309 frames in total allowing him to view the films as visualised data, stripped from narrative or dialogue but also providing a means for comparison.

In the first instance, Yannick used the machine to compare Anderson’s use of colour, directly recording the red, blue and green light used to produce each pixel of each image giving him an average colour for each frame. Yannick then arranged these averages in a two-dimensional space, highlighting, for example “the warm yellows and browns with a blue horizon, dark scenes and desaturated colours with yellow highlights,” that appear in Moonrise Kingdom. Once overlaid, the data from each film creates a colour map of-sorts drawing our attention to the strong hue and tonal similarities of Anderson’s films. We are also able to dissect when Anderson tends to use colour to create a mood, for example, by looking closer at the scenes which feature strong blue tones: this group includes the pirate attack/kidnap scenes in The Life Aquatic and Richie’s attempted suicide in The Royal Tenenbaums.

The second method is even more complex, passing the information through a neural network that is trained to recognise up to 1000 objects. The output of this experiments highlights Anderson’s common use of tents, text and textiles, screens, shelves and cars.

All of this data is presented on Yannick’s website in an impressively slick and interactive way, allowing you to roll over elements in order to see their respective scenes – well worth a look here!