1: \begin{abstract}

2: 

3: We present a method for creating inconspicuous-looking textures that, when displayed as posters in the physical world, cause visual object tracking systems to become confused.

4: As a target being visually tracked moves in front of such a poster, its adversarial texture makes the tracker lock onto it, thus allowing the target to evade.

5: This adversarial attack evaluates several optimization strategies for fooling seldom-targeted regression models: non-targeted, targeted, and a newly-coined family of guided adversarial losses.

6: Also, while we use the Expectation Over Transformation (EOT) algorithm to generate physical adversaries that fool tracking models when imaged under diverse conditions, we compare the impacts of different scene variables %conditioning variables, including viewpoint, lighting, and appearances,

7: to find practical attack setups with high resulting adversarial strength and convergence speed.

8: We further showcase that textures optimized using simulated scenes can confuse real-world tracking systems for cameras and robots.

9: %, as a means to raise awareness that innocuous patterns can harm vision-based systems simply by being present in the wild.

10: 

11: \end{abstract}

12: