Optical illusion has always attracted extensive attention, as it provides a superior self‐protection ability for both natural animals and human beings. A decade ago, this motivated the study and application of transformation optics, which provides a universal tool to manipulate light for invisibility cloaking and optical illusion. However, mainstream transformation‐optics‐based optical illusions are inherently hindered by the extreme requirements of metamaterial compositions in practice and unfavorably limited by the very large computational cost caused by their bulky state. To overcome these grand challenges, a novel and intelligent optical illusion supported by form‐free metasurfaces via a deep learning architecture is reported, which can not only render a similar illusion effect but also greatly reduces the parameter space in physics. Illustrative examples of conformal metasurfaces are presented, with a high‐fidelity inverse design from either the near‐ or far‐field in the simulation and experiment. Furthermore, a full set of intelligent systems is developed to benchmark the real‐world optical illusion applicability. The work brings the available illusion strategies closer to a wide range of in situ practical‐oriented applications and lays a foundation for the next generation of intelligent metamaterials.

The physical basis of a smart city, the wireless channel, plays an important role in coordinating functions across a variety of systems and disordered environments, with numerous applications in wireless communication. However, conventional wireless channel typically necessitates high-complexity and energy-consuming hardware, and it is hindered by lengthy and iterative optimization strategies. Here, we introduce the concept of homeostatic neuro-metasurfaces to automatically and monolithically manage wireless channel in dynamics. These neuro-metasurfaces relieve the heavy reliance on traditional radio frequency components and embrace two iconic traits: They require no iterative computation and no human participation. In doing so, we develop a flexible deep learning paradigm for the global inverse design of large-scale metasurfaces, reaching an accuracy greater than 90%. In a full perception-decision-action experiment, our concept is demonstrated through a preliminary proof-of-concept verification and an on-demand wireless channel management. Our work provides a key advance for the next generation of electromagnetic smart cities.

No abstract

A shared-information double stream hourglass network architecture for camera localisation is proposed. The contributions are two folds, first, the ordinary single stream decoder is replaced by double streams for regressing location and orientation of camera separately, and the information is shared at the end. Secondly, uncertainty estimation of loss to balance the error of location and orientation is used. The experimental results show that the proposed method achieves better performance compared with state-of-the-art methods.

No abstract