Beyond Averages: A Multi-Faceted Evaluation of Monocular Depth Estimation Using Depth-Binned Errors, Uncertainty, and Distributional Analysis

Traditional evaluation metrics for monocular depth estimation (MDE) models, such as Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), provide only a single measure of model performance, analogous to using the mean in statistics. While informative, these metrics often obscure critical details about model behavior across different depth ranges. In this paper, we propose a novel, multi-faceted evaluation framework that extends beyond averages by introducing depth-binned error analysis, uncertainty quantification, and distributional comparisons. Our method captures more comprehensive information about model performance, offering deeper insights into model reliability and accuracy. The mathematical grounding of our approach is discussed, and its efficacy is demonstrated on the NYUv2, DIODE, and TUM datasets.