Exhibition Review: Philipp Schmitt's Tunnel Vision/Declassifier
Written by Grace Russell
In his series of 35mm photographs titled Tunnel Vision, Philipp Schmitt focuses on capturing eighty main categories of objects from the Microsoft (COCO) training dataset. Using an artificial intelligence object identification program gives him the ability to test the theory that the dataset comprehensively describes and allows the program to detect everyday objects through its computer algorithm.
The Common Objects in Context (COCO) is a widely used dataset with multiple features focused on detecting, segmenting, and annotating ninety-one easily found object types. When overlaying this dataset onto photographs, it colorfully frames the detected objects and labels each with a category that fits what the program believes the object is or the category it most closely resembles.
To anyone unfamiliar with the program, it can seem to be a spectacle of sorts with its colorful frames, easy detectability, and 'precise' annotations. Still, Schmitt's project sets out to expose the myth of these intelligent machines by giving them the opportunity to run into glitches and biases found in the dataset and pinpoint exactly which data conditioned each prediction. He can, in this way, sense how the machines see and detect the many objects.
Schmitt's series consists of twenty-eight images taken throughout various areas of New York City in order to test the dataset's allegedly universal applicability in one place. Finding a few of the categories in the city proved to be challenging for him, as he mentioned that there is an unlikely chance of a cow being in Times Square or a frisbee being thrown in the brisk winter weather. The opportunity for a thorough investigation, though, provided a soaring number of objects found in more unexpected places, as opposed to finding each object in an area where it is common.
In datasets like this one, too much intricacy can confuse the computer algorithm and, according to Schmitt, keeping this in mind is essential to the process of the project. "These photos are for a computer 'audience' first. The human viewer is secondary. I shoot matter-of-fact, with little occlusion, as too much complexity will confuse the algorithm."
Using the images from "Tunnel Vision," Schmitt overlaid the COCO dataset and presented the outcome in an interactive exhibition titled "Declassifier," You can experience the exhibition online here.