| --- |
| license: apache-2.0 |
| pipeline_tag: depth-estimation |
| library_name: executorch |
| base_model: depth-anything/DA3-SMALL |
| --- |
| |
| # depth-anything-v3 |
|
|
| This repository hosts **Depth Anything 3 Small** (single-image depth estimation) |
| exported for the |
| [React Native ExecuTorch](https://www.npmjs.com/package/react-native-executorch) |
| library as ExecuTorch `.pte` programs, ready to run on device. |
|
|
| Upstream model: [depth-anything/DA3-SMALL](https://huggingface.co/depth-anything/DA3-SMALL), |
| revision `e08cab65ca0ec38e7826075418411ab90cab4da3`. |
|
|
| ## Variants |
|
|
| | Path | Backend | Precision | |
| | --- | --- | --- | |
| | `small/coreml/depth_anything_v3_small_coreml_fp16.pte` | coreml | fp16 | |
| | `small/vulkan/depth_anything_v3_small_vulkan_fp16.pte` | vulkan | fp16 | |
| | `small/xnnpack/depth_anything_v3_small_xnnpack_fp32.pte` | xnnpack | fp32 | |
| | `small/xnnpack/depth_anything_v3_small_xnnpack_int8.pte` | xnnpack | int8 | |
|
|
| ## XNNPACK int8 |
|
|
| The INT8 export uses dynamic INT8 activations and per-channel INT8 weights for |
| linear layers; convolutions remain floating point. Public inputs and outputs |
| remain float32. |
|
|
| Numerical agreement with the float reference on three photos (cats and the two |
| official DA3 SOH examples) and four upstream example illustrations: |
|
|
| | Image | Range-normalized RMSE | Correlation | |
| | --- | --- | --- | |
| | depth_cats | 0.51% | 0.999855 | |
| | SOH/000 | 0.75% | 0.999915 | |
| | SOH/010 | 0.66% | 0.999924 | |
| | demo17 | 1.50% | 0.998355 | |
| | demo18 | 0.31% | 0.999875 | |
| | demo19 | 0.81% | 0.999865 | |
| | demo20 | 1.21% | 0.999060 | |
| |
| Range-normalized RMSE is RMSE divided by the reference depth range. The worst |
| value was 1.50%, with a minimum correlation of 0.998355. These are numerical |
| export checks, not depth-accuracy benchmarks. |
| |
| ## Repository structure |
| |
| ``` |
| config.json 39 B |
| small/coreml/config.json 879 B |
| small/coreml/depth_anything_v3_small_coreml_fp16.pte 70.9 MB |
| small/vulkan/config.json 879 B |
| small/vulkan/depth_anything_v3_small_vulkan_fp16.pte 56.4 MB |
| small/xnnpack/config.json 1.5 kB |
| small/xnnpack/depth_anything_v3_small_xnnpack_fp32.pte 106.5 MB |
| small/xnnpack/depth_anything_v3_small_xnnpack_int8.pte 40.6 MB |
| ``` |
| |
| ## Compatibility |
| |
| These files are published for the **ExecuTorch v1.4.1** runtime. ExecuTorch |
| gives no forward compatibility guarantee, so an older runtime may fail to load |
| them. |
| |
| For React Native ExecuTorch model loading, see the |
| [documentation](https://docs.swmansion.com/react-native-executorch/docs/fundamentals/downloading-models). |
| To load these files in your own runtime, read the |
| [compatibility note](https://github.com/pytorch/executorch/blob/main/runtime/COMPATIBILITY.md) |
| first. |
| |
