Digital Camouflage brings that abstract research into a form almost anyone can understand.
A shirt.
A camera.
A person.
And a missing bounding box.
Why Clothing Is an Interesting Attack Surface
Clothing has a special advantage in the physical world.
It naturally moves with the human body.
A conventional adversarial patch might look suspicious if someone carried a large printed sign everywhere.
Clothing, on the other hand, is expected to cover the body.
It can contain patterns, colors, logos, graphics and textures without attracting the same attention from human observers.
That makes clothing a particularly interesting medium for researchers studying physical adversarial examples.
The 2022 CVPR research specifically explored textures that could cover arbitrary clothing shapes, allowing the adversarial pattern to become part of ordinary garments rather than appearing as an obvious standalone patch.
This is one reason Digital Camouflage works so well as an art project.
The garment itself becomes the message.
It does not announce:
"Computer-vision experiment in progress."
It simply looks like clothing.
Humans and Machines See Different Things
Perhaps the most fascinating aspect of the project is philosophical rather than technical.
Humans tend to assume that cameras and computers see what we see.
They do not.
A human looks at a person and immediately understands a three-dimensional object moving through space.
A computer-vision model receives arrays of numerical values representing pixels.
It transforms those values through layers of mathematical operations.
The final output might be:
Person — 92 percent confidence.
Car — 87 percent confidence.
Dog — 76 percent confidence.
Or, under different circumstances:
Nothing detected.
The model has no experience of a person in the human sense.
It has learned statistical relationships that allow it to produce useful predictions.
Most of the time, those predictions are remarkably effective.
But adversarial examples reveal that the relationship between visual reality and machine interpretation is not as straightforward as people sometimes assume.
A person can remain completely obvious to a human while becoming unexpectedly difficult for a particular model to classify.
That is the conceptual heart of Digital Camouflage.
The Project Is Also a Question About Trust
Weckert's work is presented as an artistic exploration rather than a guarantee of technological protection.
The project's official documentation asks viewers to consider what happens when public spaces become increasingly interpreted by machines—and whether systems that people cannot independently test should be trusted without question.
That question becomes more relevant as computer vision moves into real-world applications.
AI cameras can be used for many legitimate purposes, from industrial safety to traffic analysis and security.
But automated interpretation introduces another layer between what happens in front of a camera and what a system concludes happened.
If an AI system identifies a person, object or action incorrectly, the consequences depend on what happens next.
A detection may simply be ignored.
Or it may trigger another automated process.
The important issue is therefore not simply whether AI can see.
It is whether the AI sees accurately enough for the purpose for which it is being used.
Berlin Provides a Real-World Context
Digital Camouflage has also appeared at a moment when Berlin is expanding discussion and testing around AI-assisted video surveillance.
Berlin police currently describe an AI-supported video-protection system being configured at Kottbusser Tor. According to the official police information, the system analyzes video data automatically in real time during its technical configuration and is intended to identify certain security-relevant situations. The police state that the AI system supports human officers rather than making police decisions itself.
This context helps explain why an artwork about confusing machine vision can attract attention.
But it is important not to confuse the artist's demonstration with a test against the actual police system.
Weckert's own website explicitly says the demonstration uses a generic open-source YOLO system and does not claim that the shirt defeats any specific government surveillance system.
That distinction should remain central to any discussion of the project.
The demonstration shows a vulnerability class.
It does not establish that the garment defeats a particular surveillance deployment.
A Moving Target
There is another reason why a universal anti-camera shirt would be difficult to create.
AI systems change.
Models are updated.
Cameras improve.