Depth Anything V2 Base

Depth-Anything-V2-Base is a monocular depth estimation model that predicts a dense depth map from a single RGB image. It is the base-scale variant (ViT-B encoder) of the Depth Anything V2 family, which was accepted at NeurIPS 2024.

Key Capabilities

The model is trained on 595K synthetic labeled images and over 62 million real unlabeled images. It delivers finer-grained details and more robust predictions than its predecessor, Depth Anything V1. Compared with Stable Diffusion-based models such as Marigold and GeoWizard, Depth Anything V2 is more than 10× faster and more accurate, while being significantly lighter. The training pipeline uses a teacher model scaled to 1.3B parameters and large-scale pseudo-labeled real images to achieve strong generalization.

Performance and Evaluation

A qualitative comparison (arXiv:2406.09414, Table 1) shows that Depth Anything V2 satisfies all six desirable properties: fine detail preservation, handling of transparent objects, reflections, complex scenes, efficiency, and transferability. In the same comparison, Marigold and Depth Anything V1 each satisfy only three properties (different sets). The authors also constructed the DA-2K benchmark with precise annotations and diverse scenes to address limitations in existing test sets. Depth Anything V2 models are available from 25M to 1.3B parameters; the Base model offers a strong balance of performance and speed for deployment.