Evaluation of Monocular Depth Detectors in Eyeglasses Segmentation Scenario

Monocular depth estimation has become increasingly significant due to its wide-ranging applications in computer vision tasks, including segmentation, object recognition, and augmented reality. This study investigates the performance of several state-of-the-art monocular depth detectors in scenarios specifically focused on eyeglass segmentation. Accurate eyeglass segmentation from monocular images is a challenging but valuable task, given its implications for face recognition, virtual try-on systems, and visual effects rendering. This paper benchmarks three prominent monocular depth detectors: UniDepthV2, Depth Anything V1, and Depth Anything V2 to assess their ability to reconstruct depth maps conducive to effective segmentation of eyeglasses. The synthetic image data set was used for performance evaluation. Segmentation effectiveness was quantitatively evaluated using the Intersection over Union (IoU) metric, providing clear insight into the accuracy and reliability of each depth detector’s output. The results indicated performance variations among the models evaluated, and Depth Anything V2 demonstrated superior capability in generating accurate depth maps suitable for eyeglass segmentation, evidenced by its consistently higher IoU scores. These findings contribute valuable benchmarks that can guide the future development of monocular depth detectors, emphasizing the importance of precision in depth reconstruction tasks that involve detailed object segmentation, such as eyeglasses. Further research is suggested to explore the applicability of these detectors in real-world imaging conditions and their integration into practical segmentation pipelines.