Boyuan Sun Modi Jin Bowen Yin Qibin Hou

Abstract

We present Depth Anything at Any Condition (DepthAnything-AC), a foundationmonocular depth estimation (MDE) model capable of handling diverseenvironmental conditions. Previous foundation MDE models achieve impressiveperformance across general scenes but not perform well in complex open-worldenvironments that involve challenging conditions, such as illuminationvariations, adverse weather, and sensor-induced distortions. To overcome thechallenges of data scarcity and the inability of generating high-qualitypseudo-labels from corrupted images, we propose an unsupervised consistencyregularization finetuning paradigm that requires only a relatively small amountof unlabeled data. Furthermore, we propose the Spatial Distance Constraint toexplicitly enforce the model to learn patch-level relative relationships,resulting in clearer semantic boundaries and more accurate details.Experimental results demonstrate the zero-shot capabilities of DepthAnything-ACacross diverse benchmarks, including real-world adverse weather benchmarks,synthetic corruption benchmarks, and general benchmarks. Project Page: https://ghost233lism.github.io/depthanything-AC-page Code: https://github.com/HVision-NKU/DepthAnythingAC