Arm/rtdetr-l-onnx-raspberry

rtdetr-l optimized for Arm-based Edge Linux systems

A quantized version of rtdetr-l for object detection, exported to ONNX (.onnx) and optimized for Arm-based Edge Linux systems.

Summary

This repository contains an Arm-optimized version of rtdetr-l for object detection. RT-DETR-L is a real-time end-to-end detector that pairs a convolutional backbone with a transformer encoder-decoder and predicts a fixed set of 300 object queries without non-maximum suppression. The model is provided in ONNX (.onnx) format and run with ONNX Runtime, targeting Edge Linux systems.

Quantization is applied per region rather than uniformly: the convolutional backbone and neck use statically quantized per-channel INT8 convolutions, the transformer matrix multiplications use dynamically quantized INT8, and the prediction heads, deformable-attention sampling components and the first stem convolution are left in FP32.

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 COCO 2017 and measured performance on a representative evaluation target.

Key results

Area	Result
Model format	ONNX (.onnx)
Target device class	Edge Linux
Reference device	Raspberry Pi 5 (Cortex-A76, linux)
Primary performance result	p50 latency 673.121 ms (2.65x faster than baseline), 1.48 frames per second
Accuracy result	mAP@0.5:0.95 50.75%, mAP@0.5 69.44%
Size / memory result	38.422 MB (3.28x smaller), peak memory 390.45 MB

Original model

Field	Value
Original model	rtdetr-l
Original source	GitHub
Original developer	Ultralytics
Original model card	ultralytics/ultralytics
Original license	AGPL-3.0

Model files

File	Description
rtdetr-l_onnx_optimized.onnx	Arm-optimized 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 / accelerator	Cortex-A76, 4 cores, arm64
OS	Linux
Runtime	ONNX Runtime 1.27.0
Backend / delegate	CPU execution provider (MLAS, KleidiAI)
Batch size	1
Precision	INT8 mixed — statically quantized per-channel INT8 convolutions, dynamically quantized INT8 transformer matmuls, selected components in FP32
Runs	5000 measured runs

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
Model size (MB)	125.845	38.422	3.28x smaller
End-to-end latency p50 (ms)	1782.475	673.121	2.65x faster
End-to-end latency p90 (ms)	1805.594	692.716	2.61x faster
End-to-end latency p99 (ms)	1811.622	702.155	2.58x faster
Frames per second	0.56	1.48	2.64x
Peak memory (MB)	540.14	390.45	1.38x less
Average memory (MB)	447.11	294.78	1.52x 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	COCO 2017
Split	val2017
Number of samples	5000
Metric(s)	mAP@0.5:0.95, mAP@0.5, mAP@0.75
Evaluation runtime	ONNX Runtime 1.27.0

Accuracy results

Metric	Original / baseline	Arm-optimized	Change
mAP@0.5:0.95 (%)	50.92	50.75	-0.17 pp
mAP@0.5 (%)	69.63	69.44	-0.19 pp
mAP@0.75 (%)	54.76	54.53	-0.23 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	ONNX export of RT-DETR-L to .onnx
Quantization	Yes	PTQ hybrid — statically quantized per-channel symmetric INT8 convolutions in the backbone and neck, calibrated on 64 randomly selected COCO 2017 images; dynamically quantized INT8 transformer matmuls; nine components kept in FP32 to keep accuracy within acceptable ranges
Runtime/backend selection	Yes	ONNX Runtime CPU execution provider with MLAS and KleidiAI
Graph/runtime compatibility updates	Yes	Performed as part of the shared PT2E export pipeline, with ONNX-specific graph translation
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
Input shape	[1, 3, 640, 640]
Input type	float32
Input range	[0.0, 1.0]
Preprocessing	letterbox to 640x640 (aspect-preserving, centerd), pad value 114 (gray fill), to tensor with no mean/std normalization

Expected output

Property	Value
Output shape	[1, 300, 84]
Output type	300 object queries; 4 box coordinates (cxcywh, normalized) + 80 class scores, no objectness
Postprocessing	score = max(class_scores), sigmoid applied when the head emits logits; confidence threshold 0.001, IoU threshold 0.7, boxes converted to xyxy and mapped back to original image coordinates (letterbox padding removed and rescaled)

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

• Performance depends on the target device, runtime version, backend/delegate support, memory configuration, and system load.
• Accuracy was evaluated on COCO 2017 val2017 and may not generalize to all domains.
• This release preserves the original model's intended task and behavior, but users should validate it for their own application, data, and deployment environment.
• This repository is not a replacement for the original model documentation.

Additional notes

RT-DETR-L detects the 80 COCO object categories at a fixed 640x640 input resolution using 300 object queries, and requires no non-maximum suppression. The quantized artifact keeps the same architecture and parameter count as the source weights; only numeric precision changes.

The bundled example.py reads sample_input.jpg and writes an annotated sample_output.jpg plus a detections.json file beside itself.

• Sample input: sample_input.jpg is derived from Living room (Unsplash) by Jarosław Ceborski, via Wikimedia Commons (CC0 1.0).

About this version

Original Model: rtdetr-l by Ultralytics - 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 AGPL-3.0.

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.