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.