Evaluation of Plane Reconstruction Distortions in Monocular Depth Maps

Abstract

Monocular depth estimation is a fundamental problem in computer vision with applications in robotics, augmented reality, and 3D scene reconstruction. This study evaluates the accuracy of plane reconstruction in depth maps generated by various monocular depth estimation models. The assessment focuses on quantifying distortions in reconstructed planes using mean squared error between reconstructed in the depth image and fitted planes. Both real and synthetic images containing flat surfaces were used to systematically analyze the performance of different depth estimators. The benchmarking results provide insights into the reliability of existing monocular depth estimation methods in preserving planar structures, highlighting their strengths and limitations.