Abstract
Mirrors and reflective surfaces are common in modern environments and are important for vision-based applications, including autonomous drones and intelligent vehicle systems. Accurate mirror detection supports object recognition, semantic segmentation, scene reconstruction, and NeRF modeling. However, detecting mirrors in RGB images is challenging because they reproduce surrounding scenes rather than presenting distinct visual features. To overcome these difficulties, a deep learning-based video mirror detection approach incorporates motion randomness, enhanced edge-refinement modules, and motion cues for improved detection. The proposed MMD dataset further supports evaluation under realistic conditions, including low-light settings and scenes containing multiple mirrors. The study considers alternatives such as MMD, VMD-D, Synthetic Mirror, Real-World Mirror, Augmented Reality Mirror, Low-Light Mirror, and Dynamic Environment Mirror. Evaluation uses detection accuracy, SSIM, mirror size, spatial distribution, color contrast, lighting, and motion consistency. Weighted-sum results identify Low-Light Mirror as the highest-ranked alternative, while Dynamic Environment Mirror ranks lowest.