Review of monocular depth estimation methods

Abstract

Estimating depth from monocular images has an inherent limitation problem, which can lead to errors in understanding and recognizing the scene. Therefore, how to recover scene depth information from monocular RGB images or RGB video sequences has been a hot and challenging research topic in computer vision for decades. Traditional methods rely heavily on handcrafted features, geometric constraints, and optimization techniques. The critical approaches use in conventional monocular depth estimation include structure from motion, stereo vision, shape-from-shading, and depth from defocus. Motivated by their growing success in solving various 2D to 3D vision problems, depth estimation based on deep learning has attracted increasing interest from the community. We provide a comprehensive survey of traditional and deep learning approaches to this new and growing field of research and discuss their benefits and limitations. Reflecting on the achievements made so far, we also speculate on the future of deep learning-based stereo-depth estimation research.