Field Notes

concept

CV Dazzle

StatusStableUpdatedStands on1 sourceSupportsnotes

CV Dazzle is Adam Harvey’s computer-vision camouflage project, first made as a 2010 New York University thesis proof of concept. It uses asymmetric hair, makeup, and accessories to change the dark-light structure and geometry that a particular face detector expects.1

A detector-specific demonstration

The original looks targeted the Viola-Jones OpenCV Haarcascade detector. They aim to move a face below that detector’s threshold, which can block the later recognition stage only because no face was supplied to it in the first place. Harvey explicitly separates detection from recognition: the latter compares feature representations, and has different vulnerabilities.1

The project demonstrates a valid technical principle: computer vision is probabilistic, and a visual intervention can be effective against a named model under a named set of imaging conditions. It does not demonstrate a reusable way to defeat modern surveillance systems. The published evaluation is limited to static visible-light images, pretrained Haarcascade profiles, and comparable lighting and pose; other detectors, deep neural networks, multiple cameras, or multispectral imaging require a different strategy.1

Why it still matters

CV Dazzle is the conceptual ancestor of much anti-surveillance fashion. Its strongest lesson is methodological rather than commercial: claims should name the model, camera conditions, and test date. The later Facial recognition pipeline also shows why bypassing detection for one system is not equivalent to being unidentifiable.

  1. Adam Harvey, “CV Dazzle”, project page; local copy. The originator’s own account of the work, which fixes what the project claims and dates its claims to the detectors of its era. ↩↩↩

Built on 1 source (1 external).

Working out connections…