Analyzing the Effectiveness of Classification and Regression for Depth Estimation of Highly Dynamic Terrain

Depth estimation is considered a problem of translating RGB images to their corresponding depth maps. It plays an essential role in robotics and autonomous navigation as it allows for an understanding of the nature of the terrain. Owing to its natural importance, considerable attention has been paid by the research community in the past decade to depth estimations, including studies focusing on environmental condition-, classification-, or regression-based approaches. Among them, classification- and regression-based approaches are considered the key candidates for depth estimations, but it is not clear which approach performs better under what conditions. To investigate and define a clear method for providing an accurate depth estimation from a single RGB image, this study extensively evaluates these approaches qualitatively and quantitatively by considering a highly dynamic dataset. Based on extensive qualitative and quantitative experimental analyses, this study provides possible future directions for accurate depth estimation from a single RGB image.