Abstract

This work presents Depth Anything, a highly practical solution for robust monocular depth estimation. Without pursuing novel technical modules, the aim is to build a powerful foundation model handling images in various conditions. A data engine is designed to collect and annotate large-scale unlabeled data (∼62M), significantly enhancing data coverage and reducing generalization errors. Two strategies strengthen data scaling: data augmentation that pushes models to acquire robust visual knowledge and auxiliary supervision that inherits semantic priors from pre-trained encoders. Evaluations on six public datasets and randomly captured photos demonstrate strong generalization. Fine-tuning with NYUv2 and KITTI datasets sets new SOTA benchmarks, and improved depth models lead to superior depth-conditioned ControlNet.