Depth Anything V2 Metric Outdoor Large

Depth-Anything-V2-Metric-Outdoor-Large-hf is a monocular metric depth estimation model fine-tuned for outdoor scenes, built on the Depth Anything V2 foundation and trained on the synthetic Virtual KITTI 2 dataset. It uses the DPT architecture with a DINOv2 backbone and was trained on approximately 600,000 synthetic labeled images and 62 million unlabeled real images, achieving state-of-the-art results for both relative and absolute depth estimation.

Architecture and Training

The model is part of a family of six metric depth models (three scales for indoor and outdoor scenes). The Large variant contains 335.3 million parameters. Training follows three key practices: replacing all labeled real images with synthetic data, scaling up the teacher model capacity, and teaching student models via large-scale pseudo-labeled real images. The paper was accepted to NeurIPS 2024.

Key Strengths

• Produces finer and more robust depth predictions than Depth Anything V1.
• More than 10× faster inference than Stable Diffusion-based depth models, with higher accuracy and fewer parameters.
• Strong generalization capability across diverse outdoor scenes, demonstrated on the DA-2K evaluation benchmark.

Available Variants