Parametric Truth & Model Home
Exposé by Simone C Niquille / technoflesh Studio
As artistic-scientific doctorate project I propose:
(1) Model Home as case study of Parametric Truth, to annotate and expand with interviews the body of work described below, as comprehensive publication on the social and political consequences of computer vision and synthetic training data
(2) research and produce a film, a new chapter to Model Home, on novel synthetic data created with AI 3d model generators and the slippery slope of ground truth it introduces. What information is known to be real and true, at what weight?
The artistic-scientific research project critically examines computer vision and AI powered 3D asset generation while defining the training dataset as contemporary publishing practice. The artistic outcome, a publication and film, borrow from familiar information dissemination formats to communicate concepts that otherwise can seem abstract and distant, relating them back to the lived experience.
1 Model Home: Synthetic training data for domestic computer vision
Household robots rely on computer vision to navigate their environment. These automated domestic assistants are equipped with a camera to perceive their surroundings. But a camera does not know what it is looking at. Large datasets of 3D model files are virtually assembled into model homes as training data for these machine learning models. Model Home examines and visualizes the training datasets used to make cameras ›smart‹ to reveal the politics and hypocrisies underlying their so-called intelligence.
I have investigated domestic computer vision since 2018 and produced several film works, image essays and texts. The first work, Regarding the Pain of SpotMini (2018) is an image essay based on a YouTube video by robotics company Boston Dynamics. By rebuilding in 3D software the house navigated by the domestic robot on film, I attempt to outline the borders of what a domestic dataset might be. The film Homeschool (2019) takes place in a virtual scenography built with the data from the training dataset SceneNet. Synthetic data is used to bypass the limitations of a home’s privacy, yet, where does this data come from? Sorting Song (2021) delves into the linguistic logic underlying training dataset labelling, at the core of machine learning bias and ethical frailty. Beauty and the Beep (2024), combines research and
philosophical questions into practice by using machine learning to train a virtual chair to walk. The chair is seen searching for a place to sit while navigating the Boston Dynamics ›model home‹ initially created for Regarding the Pain of SpotMini (2018).
As publication, Model Home gathers for the first time this body of work on domestic computer vision. The aim of the publication is to provide research and process notes of these works. The technologies that are examined in the films’ narrative are also used to create the works. The publication will provide metadata and production protocol of these films and texts, while also presenting imagery that is beyond the final piece. Furthermore, through interviews with experts in related fields, Model Home will be a reference book at the intersection of media culture, computer vision, social sciences, architecture and the challenges computational optics create for cohabitation. Model Home will be the first publication of my practice and provide the space for queer principles of operation within new technologies of seeing.
2 Parametric Truth: Synthetic data and the slippery slope of ground truth
Parametric Truth is a term I coined in 2016 to describe my practice and has since been an essential lens to read contemporary visual culture. Parametric truth describes the creation and justification of truth through computational optics. It is embedded in software’s source code or produced deliberately by placing agency in digital systems. In the context of this PhD proposal, parametric truth is defined through synthetic training data for computer vision and the slippery slope of ground truth it introduces. In computer science and AI model training, ground truth is information that is known to be factual and true, a benchmark of reality.
Computer vision relies on machine learning models to extract information from a cameras field of view. Training datasets are most accurate when assembled for specific tasks. Yet, the explicit data needed, at a large enough scale, does not always exist. Think of a self driving car not recognising a deer running across the street or a tandem bike entering the road in heavy snowfall. These scenarios must exist visually for the system to learn. In cases where no visual data exists, synthetic images have acted as substitute. The video game GTA has been scraped for data to train self driving cars, but what to do in cases where no video game and no virtual world to simulate eventualities exists?
In the case of domestic computer vision models, synthetic data is necessary as the home is a private space and not enough visual data can be scraped online. 3D models are gathered from online libraries containing furniture and floor plans that are then assembled into ›homes‹ and virtually photographed. These resulting images are the synthetic image training dataset.
Yet, online 3D models are still a scarce resource limited to architectural design and game engine asset stores. Recent developments in AI generated content have turned to generating 3d models, notorious for the time and skill needed in creating them manually. In this new chapter for Model Home I will research and create work about the potential consequences of using AI generated 3D models as training data I hypothesise that it will further blur the threshold of ground truth — at what point is the training data without connection to reality, at what consequences?
Many of the mentioned technologies pass through stages of experimentation, product ideation, failure and integration into existing technologies that are incredibly opaque to map after the fact but are crucial to understanding and reading our technological realities. Parametric truth relies on annotation, documentation and interrogation in an attempt to surface software’s limits deliberately.
