Description
DepthPro2 is a sophisticated AI model from Apple designed for precise metric monocular depth estimation. This model can process various visual inputs, including single photographs, entire video sequences, and live camera feeds, generating depth information for every pixel at the input's native resolution. A key strength of DepthPro2 lies in its ability to deliver sharp depth estimations, particularly around object edges, and to maintain excellent temporal consistency, ensuring steady depth measurements from one frame to the next in video content.
Read More
The model is available in several sizes, each with its own repository and weights, allowing users to choose based on their specific needs for speed and accuracy. Larger models (L, H+) offer sharper and steadier depth results, while smaller versions (S, B) prioritize faster processing. All DepthPro2 models integrate a shared focal length estimator, `apple/DepthPro2-focal`, which automatically predicts the focal length when it is not provided, enhancing the model's versatility.
Under the hood, DepthPro2 employs a multi-view transformer architecture that jointly processes one or more frames. This transformer alternates attention mechanisms within individual frames and across frames after its initial blocks. A DPT decoder then predicts an inverse depth map for each frame by assembling tokens from various blocks into feature maps at different resolutions (1/32 to 1/4 of the input resolution). This advanced methodology enables DepthPro2 to outperform previous work in both sharpness (measured by Boundary F1 score) and temporal consistency (measured by Jitter) across datasets like Sintel and Spring.
DepthPro2 is released under the Apple Machine Learning Research Model license, which permits its use for non-commercial scientific research and academic development. This makes it a valuable resource for researchers and developers exploring advanced depth estimation techniques. The model is hosted on Hugging Face, where users can download, explore its files, and run it, with quick start examples provided for both image and video processing. It's important to note that depth is considered infinite for infinitely far scenes, such as the sky.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsHugging Face
DepthPro2 Highlights
Metric monocular depth estimation
Processes single photos, videos, and live feeds
Full resolution depth output
Sharp depth at object edges
Temporally consistent depth across frames
Available in multiple model sizes (S, B, L, H+)
Integrated focal length estimator
Multi-view transformer architecture
DPT decoder for inverse depth map prediction
Outperforms prior work in sharpness and temporal consistency
Research use only license
Command-line and live-feed examples available
Hosted on Hugging Face for easy access
Getting Started with DepthPro2
Access Model: Visit the DepthPro2 page on Hugging Face.
Select Model Size: Choose the desired model size (S, B, L, H+) based on performance needs.
Download Weights: Download the specific model weights from its repository.
Set Up Environment: Prepare your development environment with necessary dependencies.
Integrate via Code: Use the provided code examples for image or video processing.
Run Inference: Execute the model on your input data to generate depth maps.
DepthPro2's Use Cases
- Robotics Navigation
- Augmented Reality (AR)
- 3D Reconstruction
- Autonomous Vehicles
- Medical Imaging Analysis
- Video Surveillance
- Content Creation








