Physical tests found that carefully designed images could cause a depth-estimation system to miss patched objects.
The study examined how physical adversarial patches—carefully designed images intended to mislead an AI system—affect monocular depth estimation, which infers scene depth from a single camera view. In experiments on the foundation model Depth Anything, the effects were largely limited to the patch area rather than spreading across the rest of the image.
The researchers found that many learned patches contained recognizable content, such as forests or buildings, and produced corresponding depth errors. They also introduced MixCrop, an add-on training method that improved the patches’ effectiveness when their visual content was incomplete or repetitive. Physical tests showed that objects covered by the patches could effectively vanish from the model’s predicted depth map.
What the patches changed
The researchers developed a general-purpose pipeline for attacking monocular depth-estimation systems with physical adversarial patches. In tests with Depth Anything, the patches’ effects were largely confined to the covered area, suggesting greater resistance to wider image disruption than in traditional models. The patches often contained meaningful visual patterns, including forests or buildings, that were linked to the incorrect depth predictions they produced.
The researchers’ MixCrop method improved attack effectiveness, particularly when patch content was incomplete or repetitive. Experiments, including physical attacks, found that patched objects could effectively disappear from the predicted depth map. The patches also generalized across different objects and scenes.
Why depth errors matter
Depth-estimation systems can help machines interpret their surroundings for applications such as autonomous driving and robotic navigation. The findings show that physical patches may cause one such system to miss the depth of an object, identifying a potential weakness that should be considered when assessing the safety and reliability of these applications.
The study also indicates that the attacks are not necessarily broad failures across an entire image: their main effects were concentrated around the patch. That distinction may help researchers understand how these systems respond to misleading visual content and how defenses should be tested.
Evidence and caveats
The conclusions are based on experiments with the Depth Anything monocular depth-estimation foundation model, including tests using physical patches. The abstract reports experiments across objects and scenes, but it does not provide sample sizes, attack success rates, testing conditions or comparisons with specific other models. The findings therefore establish a demonstrated vulnerability in the studied system, rather than showing that all depth-estimation systems or deployed vehicles and robots would respond in the same way.
// Source
Scientific Reports · 2026 · DOI: 10.1038/s41598-026-76240-2
Authors: Yuan Zhang, Jian Wang, Jing Yin
Institutions: Xi'an Jiaotong University, Xinjiang University, China Telecom