A survey of monocular depth estimation in non-standard scenarios： From environmental degradation to spatial distortion

Monocular depth estimation was regarded as a key visual perception technology in fields such as autonomous driving， robotic environment perception and virtual reality， by which the geometric mapping from two-dimensional images to three-dimensional physical space was supported. However， traditional methods mostly relied on ideal imaging conditions and standard perspective projection assumptions， and their performance was significantly degraded in non‑standard scenarios. First， the frontiers of research on monocular depth estimation in non‑standard scenarios were systematically reviewed， and the core technical challenges were distilled into two major aspects： environmental degradation and spatial distortion. To address the image degradation caused by complex environments such as rain， fog， and nighttime， optimization strategies for improving model robustness were summarized from three aspects： generative augmentation and foundation model priors， self‑supervised cross‑domain feature adaptation， and explicit physical constraints with multimodal fusion. To address the spherical spatial distortion caused by the equirectangular projection （ERP） in panoramic fields of view， novel depth estimation paradigms that adapt to spherical geometric properties were analyzed along three technical paths： supervised learning， self‑supervised learning， and semi‑supervised learning with knowledge distillation. Second， it was pointed out that cross‑domain transfer based on foundation model priors was a common technical trend for solving the challenges of data scarcity and distribution shift in the above two types of scenarios. Finally， a prospect for constructing a general monocular depth estimation method was presented from four dimensions： unified prior transfer driven by foundation models， composite perception of environment and space， multimodal heterogeneous fusion， and standardization of evaluation benchmarks.