Depth Anything, published in January 2024 by Lihe Yang and co-authors and accepted at CVPR 2024, is a robust solution for monocular depth estimation - guessing how far away every pixel is from a single ordinary photograph, with no stereo camera or depth sensor. Recovering 3D distance from one 2D image is fundamentally ambiguous, so the practical challenge is generalization: working on arbitrary scenes the model has never seen.
The paper’s contribution is data, not architecture. The team built a data engine that collected and automatically annotated roughly 62 million unlabeled images, using a teacher model to produce pseudo-depth labels and then training a student on that vast, varied corpus. The scale and diversity of the data, rather than any novel network design, gave the model strong zero-shot generalization across six public benchmarks and real-world photos without fine-tuning.
It built on the frozen-feature philosophy of DINOv2-style backbones and quickly became a default tool for grab-and-go depth, with a Depth Anything V2 following.
Why business readers should care: cheap, reliable depth from a single camera unlocks 3D understanding without expensive LiDAR - useful in AR, robotics, photography effects like portrait blur, and any application that needs to know scene geometry from commodity cameras. Depth Anything is another case where a large auto-labeled dataset, not a clever model, delivered the leap.