| --- |
| tags: |
| - depth-estimation |
| library_name: coreml |
| license: apache-2.0 |
| --- |
| |
| # Depth Anything V2 Core ML Models |
|
|
| Depth Anything V2 was introduced in [the paper of the same name](https://arxiv.org/abs/2406.09414) 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](https://arxiv.org/abs/2401.10891) by Lihe Yang et al., and was first released in [this repository](https://github.com/LiheYoung/Depth-Anything). |
|
|
| ## Model description |
|
|
| Depth Anything V2 leverages the [DPT](https://huggingface.co/docs/transformers/model_doc/dpt) architecture with a [DINOv2](https://huggingface.co/docs/transformers/model_doc/dinov2) backbone. |
|
|
| 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. |
|
|
| <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/depth_anything_overview.jpg" |
| alt="drawing" width="600"/> |
|
|
| <small> Depth Anything overview. Taken from the <a href="https://arxiv.org/abs/2401.10891">original paper</a>.</small> |
|
|
| ## Evaluation - Variants |
|
|
| | Variant | Parameters | Size (MB) | Weight precision | Act. precision | abs-rel error | abs-rel reference | |
| | ------------------------------------------------------- | ---------: | --------: | ---------------- | -------------- | ------------: | ----------------: | |
| | [small-original](https://huggingface.co/pcuenq/Depth-Anything-V2-Small-hf) (PyTorch) | 24.8M | 99.2 | Float32 | Float32 | | | |
| | [DepthAnythingV2SmallF32](DepthAnythingV2SmallF32.mlpackage) | 24.8M | 99.2 | Float32 | Float32 | 0.0072 | small-original | |
| | [DepthAnythingV2SmallF16](DepthAnythingV2SmallF16.mlpackage) | 24.8M | 49.8 | Float16 | Float16 | 0.0089 | small-original | |
|
|
| 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. |
|
|
| | Device | OS | Inference time (ms) | Dominant compute unit | |
| | -------------------- | ---- | ------------------: | --------------------- | |
| | iPhone 12 Pro Max | 18.0 | 31.10 | Neural Engine | |
| | iPhone 15 Pro Max | 17.4 | 33.90 | Neural Engine | |
| | MacBook Pro (M1 Max) | 15.0 | 32.80 | Neural Engine | |
| | MacBook Pro (M3 Max) | 15.0 | 24.58 | Neural Engine | |
|
|
|
|
| ## Download |
|
|
| Install `huggingface-cli` |
|
|
| ```bash |
| brew install huggingface-cli |
| ``` |
|
|
| To download one of the `.mlpackage` folders to the `models` directory: |
|
|
| ```bash |
| 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. |
|
|
| ## Integrate in Swift apps |
|
|
| The [`huggingface/coreml-examples`](https://github.com/huggingface/coreml-examples/blob/main/depth-anything-example/README.md) repository contains sample Swift code for `DepthAnythingV2SmallF16.mlpackage` and other models. See [the instructions there](https://github.com/huggingface/coreml-examples/tree/main/depth-anything-example) to build the demo app, which shows how to use the model in your own Swift apps. |
|
|
