Depth Anything Base

LiheYoung/depth-anything-base-hf is a monocular depth estimation model that combines a DINOv2 backbone with the DPT architecture to produce high-quality depth maps from a single RGB image.

Architecture and training

The model leverages a Vision Transformer (DINOv2) encoder and a DPT decoder, a design that yields up to 28% relative improvement over fully convolutional networks for depth estimation. It was trained on a combination of 1.5 million labeled images and over 62 million unlabeled images using a teacher-student pipeline, where the teacher generates pseudo-labels and the student is trained with strong augmentations to improve robustness.

Performance

Depth Anything achieves state-of-the-art results on both relative and absolute depth estimation benchmarks, demonstrating strong generalization across diverse scenes and lighting conditions. The paper was accepted at CVPR 2024.

This model is hosted as a managed, OpenAI-compatible API on gigarouter, allowing developers to integrate monocular depth estimation into their applications without managing infrastructure.