When Clothes Confuse AI: What an “Invisible” T-Shirt Really Tells Us About AI Cameras

How clothing patterns, movement and real-world conditions can affect computer vision

Introduction

‍

A person walks in front of an AI camera. The person is clearly visible to us, yet the system may not always identify the person with the same confidence. This sounds strange until we understand one basic fact: a camera records an image, while an AI model interprets that image through patterns learned during training.

This idea became popular through demonstrations such as “The T-Shirt Invisibility Cloak,” where specially designed clothing was shown as a way to interfere with person-detection systems. The wearer does not become physically invisible. The camera still captures the person. Instead, the visual pattern can make a particular AI model less confident or cause an incorrect prediction.

So, how can a piece of clothing affect machine vision? And what does this tell us about the strengths and limitations of AI surveillance?

‍

How Does an AI Camera See?

‍

A normal CCTV camera mainly captures and records video. An AI camera adds software that analyses the video for a specific task, such as detecting a person, vehicle, face or event. 

A simple way to understand the process is:

‍

‍

The model does not understand a person exactly as a human does. It processes numerical representations learned from training data. For a person detector, the question is closer to “Do these visual patterns match the class ‘person’?” than “I know this is a human.”

‍

Think About It

‍
A human sees a person wearing a complex shirt and easily separates the shirt design from the person. An AI model must decide what the visual patterns in the image mean for its particular task.

‍

When Clothing Becomes Difficult

‍

‍

Clothing can become challenging when it changes the visual information available to a model.

Complex patterns can add many edges, repeated shapes and textures to an image. These do not automatically confuse an AI camera, but they can influence the features a detector uses.

Low contrast can create another problem. If clothing is close in colour or brightness to the background, the boundary between the person and the environment may be harder to separate, especially in poor lighting or low-resolution footage.

Movement creates another challenge. A loose jacket, flowing fabric or layered clothing changes shape as the person walks or turns. Researchers designing adversarial T-shirts have had to account for these non-rigid changes because a printed pattern does not stay flat on a moving body.

‍

The T-Shirt Invisibility Cloak

‍

‍

The video “The T-Shirt Invisibility Cloak” is a simple introduction to a larger research area called physical adversarial examples.

In adversarial machine learning, researchers study inputs deliberately designed to make a model produce an incorrect prediction. With clothing, the visual pattern is physically printed on a garment and then captured by a camera.

A 2019 study by Xu and colleagues demonstrated an adversarial T-shirt designed to interfere with person detectors in the physical world. Under their test conditions, the reported physical attack success rate against YOLOv2 was 57%. The work also modelled cloth deformation caused by movement.

A 2022 CVPR study on adversarial texture extended the idea to different viewing angles and tested physical clothing such as T-shirts, skirts and dresses.

But there is an important limitation: an adversarial garment is not a universal invisibility cloak. Different cameras and AI systems use different models, training data and processing pipelines. A pattern that affects one detector may have little effect on another.

‍

Why Does the AI Get It Wrong?

‍

Clothing may be only one part of the problem. A model can receive features different from what it learned during training. Camera angle, distance, movement, lighting, motion blur, compression and occlusion can also change the image.

Imagine the same person in four situations:

‍

‍

There may not be one fixed answer. AI performance depends on the combination of conditions.

‍

AI Camera vs Human Vision

‍

Humans use context. We understand that clothes wrinkle, people turn, shadows change and objects may temporarily block part of a body. An AI detector is more task-specific and can become less reliable when input conditions differ from the data used to train or evaluate it.

This does not make AI useless. It makes realistic testing important.

‍

Building More Reliable AI Surveillance

‍

The lesson is not that AI cameras should be abandoned. They should be tested under difficult, realistic conditions.

Testing can include different clothing patterns, colours, body positions, distances, camera angles, lighting conditions and crowded scenes. Security teams should also distinguish between person detection, facial recognition, tracking and behaviour analysis because these are different tasks. ^[6]

Most importantly, an automated alert should be treated as a prediction, not unquestionable truth. Important decisions should include appropriate human review.

‍

A New Research Direction

‍

Research is also moving beyond ordinary visible-light cameras. A 2026 CVPR paper explored adversarial clothing designed to affect both visible and infrared surveillance systems. At the same time, researchers continue to study ways of making models more robust against physical adversarial attacks.

This creates a continuing security cycle:

‍

‍

That cycle is a normal part of security research.

‍

‍

At a Glance: Human Vision vs AI Detection

‍

‍

Conclusion

‍

The idea of an “invisible T-shirt” is fascinating because it challenges a common assumption about AI: if a machine can see an image, we expect it to understand that image correctly.

A person can remain completely visible while a particular AI detector becomes less confident or makes a wrong prediction. Clothing patterns, texture, folds, movement, lighting, camera angle and other factors can interact with a model in unexpected ways. Research has demonstrated physical adversarial clothing against person detectors, while newer work is exploring more complex surveillance conditions.

The real lesson is not that a shirt can make someone invisible. AI vision is a prediction system with strengths and weaknesses. Understanding those weaknesses helps researchers build better defences, helps security teams evaluate systems realistically, and reminds us that automated surveillance should be deployed with technical care, human oversight and respect for privacy.

‍

References

• The T-Shirt Invisibility Cloak
• YouTube video referenced for the article’s introductory example: https://www.youtube.com/watch?v=NyofHyRm5CQ
• Xu, K., Zhang, G., Liu, S., Fan, Q., Sun, M., Chen, H., Chen, P.-Y., Wang, Y., & Lin, X. (2020). Adversarial T-shirt! Evading Person Detectors in a Physical World. Computer Vision – ECCV 2020, 665–681. https://doi.org/10.1007/978-3-030-58558-7_39
• Study demonstrating a physical adversarial T-shirt against person detection and modelling non-rigid cloth deformation during movement. https://arxiv.org/abs/1910.11099
• Thys, S., Van Ranst, W., & Goedemé, T. (2019). Fooling Automated Surveillance Cameras: Adversarial Patches to Attack Person Detection. CVPR Workshops.
• Earlier work showing physical adversarial patches can reduce the accuracy of person detectors. https://openaccess.thecvf.com/content_CVPRW_2019/html/CV-COPS/Thys_Fooling_Automated_Surveillance_Cameras_Adversarial_Patches_to_Attack_Person_Detection_CVPRW_2019_paper.html
• Hu, Z., Huang, S., Zhu, X., Sun, F., Zhang, B., & Hu, X. (2022). Adversarial Texture for Fooling Person Detectors in the Physical World. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 13307–13316.
• Study extending physical adversarial clothing to multiple viewing angles and garments including T-shirts, skirts and dresses. https://openaccess.thecvf.com/content/CVPR2022/html/Hu_Adversarial_Texture_for_Fooling_Person_Detectors_in_the_Physical_World_CVPR_2022_paper.html
• Long, J., Jiang, T., Liu, H., Ma, C., Zhou, W., Yang, Y., & Yao, W. (2026). Thermally Activated Dual-Modal Adversarial Clothing against AI Surveillance Systems. CVPR 2026.
• Recent work exploring adversarial clothing against visible and infrared surveillance systems. https://openaccess.thecvf.com/content/CVPR2026/html/Long_Thermally_Activated_Dual-Modal_Adversarial_Clothing_against_AI_Surveillance_Systems_CVPR_2026_paper.html
• Li, S., Zhang, S., Chen, G., Wang, D., Feng, P., Wang, J., Liu, A., Yi, X., & Liu, X. (2023). Towards Benchmarking and Assessing Visual Naturalness of Physical World Adversarial Attacks. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 12324–12333.
• Useful context on physical adversarial attacks and the challenge of making them visually natural. https://openaccess.thecvf.com/content/CVPR2023/papers/Li_Towards_Benchmarking_and_Assessing_Visual_Naturalness_of_Physical_World_Adversarial_CVPR_2023_paper.pdf ^[6]

‍

‍