Spotter #Artis
year: 2018 - 2022
material/technique: robot-camera, screens, audio & HD video/Machine Vision, Machine Learning, Object Detection, GAN
artists-in-residence: Machine Wilderness, ARTIS Zoo, Amsterdam
organisation: Zone2Source/FoAM/ARTIS
online info: website Machine Wilderness
For the Machine Wilderness artist-in-residence at ARTIS Zoo, the artists created an autonomous spotter camera that recognizes, follows and records a specific animal. Inspired by his observations Spotter tries to make himself a representation of the animal. This learning process starts with noise but gradually we see dreamy representations of the animal emerge. This AI project adds a contemporary approach to ARTIS's long tradition of artists observing and portraying animals: Natura Artis Magistra (Nature is the teacher of the arts). The artists have set up Spotter at various animal enclosures: meerkat, trumpeter swan, tern & ibis, ibex, mandrill, giraffe, zebra, weaverbird. The HD animations of the respective learning processes have a duration of about 90 minutes each.
A pan-tilt camera is equipped with real-time object detection software (YOLO). The robot looks around and when it detects the animal, the camera zooms in to take close-up pictures. During the day, it collects thousands of images that serve as input for the generation process. The GAN (Generative Adversarial Network) is like a game played between an artist (generator) and a critic (discriminator). The artist learns to create images of the animals, based on the feedback from the critic. Only the critic has access to the pictures. At first, both the artist and the critic have no idea what the animal looks like. The critic learns to detect patterns in the pictures, and the artist creates a first random image. The critic provides detailed feedback on shape, texture, and color characteristics and informs the artist how to improve the image. The artist creates a subsequent image based on this feedback in hopes of getting a better rating. Meanwhile, the critic has discovered more patterns in the pictures and incorporates this knowledge into his new feedback, and so on. This iterative process leads to a continuous stream of pictures, that have been captured frame by frame in long animations. A similar approach was used to generate sound, based on recordings of the typical sounds of the animal.
The development of Spotter is made possible by the support of the Mondriaan Fund.
|