Coreml Depth Anything V2 Small

🚀 Depth Anything V2 Core ML Models

Depth Anything V2 uses synthetic data and a larger teacher model for finer and more robust depth predictions, based on the architecture of the original Depth Anything.

🚀 Quick Start

To get started with Depth Anything V2 Core ML Models, you first need to download the model. Here are the steps:

Installation

Install huggingface-cli:

brew install huggingface-cli

Downloading the Model

To download one of the .mlpackage folders to the models directory:

huggingface-cli download \
  --local-dir models --local-dir-use-symlinks False \
  apple/coreml-depth-anything-v2-small \
  --include "DepthAnythingV2SmallF16.mlpackage/*"

To download everything, skip the --include argument.

✨ Features

• Advanced Architecture: Depth Anything V2 leverages the DPT architecture with a DINOv2 backbone.
• Large-scale Training: Trained on ~600K synthetic labeled images and ~62 million real unlabeled images, achieving state-of-the-art results for both relative and absolute depth estimation.

📦 Installation

As mentioned above, you need to install huggingface-cli first:

brew install huggingface-cli

Then you can download the model according to your needs.

💻 Usage Examples

Integrate in Swift apps

The huggingface/coreml-examples repository contains sample Swift code for DepthAnythingV2SmallF16.mlpackage and other models. See the instructions there to build the demo app, which shows how to use the model in your own Swift apps.

📚 Documentation

Model description

Depth Anything V2 was introduced in the paper of the same name by Lihe Yang et al. It uses the same architecture as the original Depth Anything release, but uses synthetic data and a larger capacity teacher model to achieve much finer and robust depth predictions. The original Depth Anything model was introduced in the paper Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data by Lihe Yang et al., and was first released in this repository.

The model is trained on ~600K synthetic labeled images and ~62 million real unlabeled images, obtaining state-of-the-art results for both relative and absolute depth estimation.

Depth Anything overview. Taken from the original paper.

Evaluation - Variants

Evaluated on 512 landscape images from the COCO dataset with aspect ratio similar to 4:3. Images were streched to a fixed size of 518x396, and the groundtruth corresponds to the results from the PyTorch model running on CUDA with float32 precision.

Evaluation - Inference time

The following results use the small-float16 variant.

📄 License

This project is licensed under the apache-2.0 license.