| --- |
| language: |
| - en |
| - zh |
| tags: |
| - coreml |
| - depth-estimation |
| - panoramic |
| - ane |
| - apple-neural-engine |
| - quantized |
| - int8 |
| - ios |
| license: apache-2.0 |
| pipeline_tag: depth-estimation |
| --- |
| |
| # DAP (Depth Anything Panoramic) - 256x512 INT8 CoreML Model |
|
|
| This repository contains the quantized CoreML model for **Depth Anything Panoramic (DAP)**, tailored and optimized for Apple Neural Engine (ANE) on iOS devices (e.g. iPhone 15 / iPhone 16 series). |
|
|
| ## Model Overview |
|
|
| - **Architecture**: Depth Anything V2 Panoramic Backbone (DAP) |
| - **Input Resolution**: `256 x 512` (Equirectangular Panoramic Frame, RGB, 3 channels) |
| - **Quantization**: INT8 Linear Weight Quantization |
| - **Runtime Acceleration**: Apple Neural Engine (ANE) + MPSGraph |
| - **Package Format**: Apple CoreML `.mlpackage` |
| - **Compressed Archive**: `dap_256x512_int8.mlpackage.zip` (277 MB) |
|
|
| ## Benchmark Performance on Hardware (iPhone 15) |
|
|
| | Metric | Measured Value | |
| | :--- | :--- | |
| | **Engine / Compute Unit** | Apple Neural Engine (`.all`) | |
| | **Average Inference Latency** | **82.5 ms** (~12.1 FPS) | |
| | **Cold Start / Load Latency** | ~650 ms | |
| | **Runtime Memory Peak** | **410 MB** | |
| | **Output Metric** | Relative / Metric Depth Map (`256 x 512`, FP32) | |
|
|
| ## Quickstart (Download & Extract) |
|
|
| You can download and extract the model into your WalkMate project or custom iOS application using: |
|
|
| ```bash |
| # Download |
| curl -L -o dap_256x512_int8.mlpackage.zip https://huggingface.co/Sysking/dap-quantized/resolve/main/dap_256x512_int8.mlpackage.zip |
| |
| # Extract |
| unzip dap_256x512_int8.mlpackage.zip |
| ``` |
|
|
