UltraFace RFB-320 optimized for Arm-based Edge Linux devices
An INT8-quantized version of ultraface-rfb-320 for object detection, exported to ONNX Runtime (.onnx) and optimized for Arm-based Edge Linux devices.
Summary
This repository contains an Arm-optimized version of ultraface-rfb-320 for object detection, quantized to INT8 via static per-channel post-training quantization. The model is provided in ONNX Runtime (.onnx) format, targeting Edge Linux systems.
This version is intended to demonstrate efficient inference on Arm-based platforms while preserving the original model's intended behavior. Arm has evaluated this model on WIDER FACE and measured performance on a representative evaluation target.
Key results
| Area | Result |
|---|---|
| Model format | ONNX Runtime (.onnx) |
| Target device class | Edge Linux |
| Reference device | Raspberry Pi 5 (Cortex-A76, linux Raspberry Pi OS Trixie) |
| Primary performance result | 6.66 ms p50 latency (1.99x vs baseline) |
| Accuracy result | WIDER FACE AP Hard 43.17% (-0.24 pp vs baseline) |
| Size / memory result | 0.761 MB (1.41x smaller), peak memory 62.56 MB |
Original model
| Field | Value |
|---|---|
| Original model | ultraface-rfb-320 |
| Original source | GitHub |
| Original developer | Linzaer |
| Original model card | Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB |
| Original license | MIT |
Model files
| File | Description |
|---|---|
ultraface-rfb-320_raspberry_onnx_optimized.onnx |
Arm-optimized INT8 model for deployment |
example.py |
Minimal inference example |
pyproject.toml |
Pinned runtime dependencies for example.py, resolved with uv |
uv.lock |
Locked dependency resolution for pyproject.toml |
config.yaml |
Model I/O contract used by the example |
benchmarks/ |
FP32 baseline and Arm-optimized benchmark records |
Performance
Performance was measured on the reference configuration below. Results are intended to make the optimization reproducible but do not guarantee identical performance on every Arm-based system.
Reference configuration
| Field | Value |
|---|---|
| Device / platform | Raspberry Pi 5 |
| CPU | Cortex-A76, 4 cores, 2.4 GHz |
| OS | linux, Raspberry Pi OS Trixie |
| Runtime | onnxruntime 1.26.0 |
| Execution backend | cpu (MLAS, KleidiAI) |
| Precision | INT8, static, symmetric, per-channel weights |
| Batch size | 1 |
| Input resolution | 320x240 |
| Runs / warmup | 100 / 10 |
Performance results
| Metric | Original / baseline | Arm-optimized | Improvement |
|---|---|---|---|
| Model size (MB) | 1.070 | 0.761 | 1.41x smaller |
| End-to-end latency p50 (ms) | 13.262 | 6.66 | 1.99x faster |
| End-to-end latency p90 (ms) | 13.314 | 6.703 | 1.99x faster |
| Frames per second | 75.4 | 150.16 | 1.99x |
| Peak memory (MB) | 65.59 | 62.56 | 1.05x less |
| Average memory (MB) | 65.58 | 62.56 | 1.05x less |
Accuracy
Accuracy was evaluated using the same preprocessing, input resolution, and evaluation protocol described below. Where possible, the optimized model is compared against the original model under the same evaluation conditions.
Evaluation setup
| Field | Value |
|---|---|
| Dataset | widerface |
| Sample count | 3226 |
| Metric(s) | WIDER FACE AP Easy, WIDER FACE AP Medium, WIDER FACE AP Hard |
| Runtime | onnxruntime 1.26.0 (cpu) |
Accuracy results
| Metric | Original / baseline | Arm-optimized | Change |
|---|---|---|---|
| WIDER FACE AP Easy (%) | 78.53 | 78.38 | -0.15 pp |
| WIDER FACE AP Medium (%) | 69.49 | 69.21 | -0.28 pp |
| WIDER FACE AP Hard (%) | 43.41 | 43.17 | -0.24 pp |
Accuracy was measured using the evaluation setup described above. Users should re-evaluate the model on their own data before production use.
Arm optimization approach
Arm optimized this model for efficient inference on Arm-based platforms using a hardware-aware conversion and validation flow.
For this release, Arm used:
| Optimization area | Applied? | Notes |
|---|---|---|
| Model conversion | Yes | Converted to ONNX Runtime .onnx |
| Quantization | Yes | INT8 PTQ-static — symmetric, per-channel weights; calibrated on 2000 widerface samples |
| Runtime/backend selection | Yes | ONNX Runtime CPU execution provider |
| Graph/runtime compatibility updates | Yes | Performed as part of the onnxruntime export pipeline |
| Accuracy validation | Yes | Compared against the original model or published baseline |
| Performance validation | Yes | Measured on the reference Arm platform |
The goal of this process is to improve deployment characteristics such as latency, memory use, model size, and runtime compatibility while preserving the model's intended behavior. Detailed conversion scripts, calibration configuration, or backend-specific implementation details may be provided separately where appropriate.
Using this model
Install dependencies
Dependencies are declared in pyproject.toml, which ships with this repository. Resolve and install them into a local virtual environment with uv:
uv python install
uv sync --frozen
Run the example
uv run example.py
Expected input
| Property | Value |
|---|---|
| Shape | [1, 3, 240, 320] |
| Dtype | float32 |
| Range | -0.992 to 1.0, after normalization |
| Preprocessing | resize to 320x240 (anisotropic stretch, no letterbox padding), then normalize as (x * 255 - 127) / 128 |
Expected output
| Property | Value |
|---|---|
| Shape | confidences [1, 4420, 2], locations [1, 4420, 4] |
| Format | raw SSD heads: no softmax, no prior decode, no NMS |
| Postprocessing | softmax over the 2 classes, SSD center-form decode against 4420 cached priors (center_variance 0.1, size_variance 0.2), confidence threshold 0.1, NMS at IoU threshold 0.3, boxes in xyxy format mapped back to original image coordinates |
Intended use
This model is intended for developers evaluating object detection workloads on Arm-based platforms. It is suitable as a reference implementation for benchmarking, prototyping, and integration exploration.
Limitations
- Evaluated on the WIDER FACE validation set (3226 images).
- Fixed 320x240 input resolution; the model and its anchor priors were trained and calibrated for this exact size.
- Performance measurements are from Raspberry Pi 5 hardware; results may differ on other Arm devices.
Additional notes
- The full network, including the classification and regression detection heads, is quantized to INT8 with no layers excluded — the exported graph requantizes head outputs to INT8 before the final concatenation regardless, so excluding them would not preserve additional precision.
- The sample input image is by Sandra Gabriel, used under the Unsplash License (free to use, no attribution required): https://unsplash.com/photos/a-group-of-people-standing-next-to-each-other-HbLqjflGgXw
About this version
Original Model: UltraFace RFB-320 by Linzaer - Repository
Optimization/conversion: Arm-Optimized version for execution on Arm-based platforms.
Converted/optimized by: Arm
License: The Original Model and the Optimized Model are subject to MIT.
This repository contains a converted or optimized version of the Original Model (the “Optimized Model”). The Original Model has been converted or optimized as described above for execution on Arm-based platforms.
No retraining or fine-tuning of the Original Model was performed as part of the conversion or optimization. The conversion or optimization was not intended to change the Original Model’s behavior or intended use.
Original Model and Documentation
For information about the Original Model, including its development, training data, intended uses, limitations and other relevant information, please refer to the Original Model repository. Information in that repository was provided by the original developer or other third parties and, unless expressly stated otherwise, has not been independently verified by Arm.
Licenses and Third-Party Terms
Use of the Original Model and the Optimized Model is subject to the applicable licenses, usage restrictions and other terms identified above and in the relevant repositories. Publication of the Optimized Model does not grant any rights beyond those provided under the applicable license terms.
You are responsible for reviewing those terms and ensuring that your use of the Original Model and the Optimized Model is permitted.
Purpose of this Release
The Optimized Model is provided as a reference implementation to demonstrate and evaluate execution and performance on Arm-based systems. It is not a production-ready or supported solution.
Arm’s publication of the Optimized Model does not constitute an endorsement or certification of the Original Model or a representation that the Optimized Model is suitable for production use or any particular purpose.
To the fullest extent permitted by applicable law (i) the Optimized Model is provided “as is.” Arm makes no representations or warranties that the Original Model, the Optimized Model or their outputs are accurate, safe, secure, non-infringing, legally compliant, suitable for production use or fit for any particular purpose; and (ii) Arm will not be liable for any loss or damage arising from or in connection with the Optimized Model, its use or its outputs.
You are responsible for independently evaluating the Optimized Model, its outputs and its suitability for your intended use, including compliance with applicable legal, regulatory, safety and security requirements.
Arm does not commit to provide ongoing support, maintenance or updates for the Optimized Model. Any use of or reliance on the Optimized Model or its outputs is at your own risk.