Roundabouts: rotoscopy, video and phases

Using images from an earlier project (Rotondes. Panorames i fases per a un espectador en moviment, Marcel Pié / Estampa, 2017), a neural net was trained to rotoscope, in other words, to convert video footage into drawing. Training was also done in the opposite direction: a photographic shot was generated from a drawn image. As the training pairs of images were video footage and rotoscopes of various roundabouts in the Ebro Delta area, the type of photorealistic image generated by the network is connected with this environment (highways, traffic signs, flat landscape, plant life and the sky).

Using these tools, we sought mechanisms to make the network generate imaginary landscapes. The strategy was to provide the network with just fragments of the rotoscoped landscape: faced with incomplete inputs, the network generated an image that lies somewhere between photorealism and the glitch.

Rotondes. Panorames i fases per a un espectador en moviment (euronal network version)

Diptych where you can see the input (left) and the output generated by the network (right):

Recursive conversion: original image in video rotoscopied by a neural network and converted into video again through a neural network: