Exploring the Benefits of Ensemble Techniques Involving MiDaS and MonoDepth2 Models for Monocular Depth Estimation

Over the past decades, research on monocular depth estimation has witnessed significant progress and has garnered substantial attention. This technique holds immense potential due to its broad range of applications. However, the complexity involved in generating accurate depth maps from a single image of real-world scenarios has made monocular depth estimation a thriving field for computer vision enthusiasts. In this paper, our focus revolves around two benchmark models, namely MonoDepth2 and MiDaS, which are widely recognized for their proficiency in depth estimation. These models serve as the foundation of our research, as we explore their capabilities in generating depth maps. It is worth noting that we employ pretrained versions of these models, without subjecting them to fine-tuning, to produce depth maps from real-world image inputs. Our dataset includes not only regular scenarios but also incorporates adversarial cases, adding further complexity to the task. We employ a late fusion technique to enhance our results' accuracy and reliability. This involves to ensemble of the outputs of both MonoDepth2 and MiDaS at various stages during the decision-making process. By integrating the strengths of these models, our goal is to improve the overall performance and robustness of our depth estimation system. Furthermore, we present a detailed comparative analysis based on the output produced by our proposed approach. This work offers an understanding of the advantages and disadvantages of each model, their performance under different scenarios, and their compatibility with real-world image inputs. We present a thorough comprehension of the capabilities and limitations of MonoDepth2 and MiDaS, enabling researchers and practitioners to make wise choices when applying these models in research.