CMS Special Seminar

This presentation explores solutions to the limitations of RGB-based perception in visually challenging scenarios. We approach this issue from two angles: (1) when external environmental conditions impair visibility, and (2) when the objects of interest possess properties that inherently hinder perception. The first case involves scenarios such as drastic lighting changes, strong sunlight, or complete darkness. To ensure consistent perception across such conditions, we utilize thermal infrared (TIR) imaging, which supports robust performance across a wide range of computer vision applications. The second case focuses on objects like transparent materials, which are often undetectable by conventional RGB or depth sensors. To address this, we adopt recent advances in generative models—specifically diffusion models and Gaussian Splatting—to recover the 3D structure of transparent objects using only RGB input.