Arm/efficient-sam-s-int8-mlas-onnx

EfficientSAM-S optimized for Arm-based Cloud CPU

EfficientSAM-S box-prompted image segmentation optimized as an INT8 ONNX Runtime (.onnx) model for Arm-based Cloud CPU systems.

Summary

This repository contains an Arm-optimized version of EfficientSAM-S for image segmentation, with INT8 weights (8-bit, per-channel, symmetric) and dynamic INT8 activations. The model is provided in ONNX Runtime (.onnx) format, targeting Cloud CPU 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 COCO 2017 and measured performance on a representative evaluation target.

View the model listing on the Arm AI Portal for full data and code examples.

Key results

Area	Result
Model format	ONNX Runtime (.onnx)
Target device class	Cloud CPU
Reference device	AWS Graviton G4 (Neoverse-V2, Linux Ubuntu 24.04.4 LTS)
Primary performance result	971.796 ms p50 latency (1.03 FPS), 1.70x faster than FP32
Accuracy result	mIoU 74.71% (ground-truth box prompts); Segm mAP@0.5:0.95 42.91% (ViTDet-H detector prompts)
Size / memory result	31.818 MB (.onnx), 3.36x smaller than FP32; 1556.53 MB peak memory

Original model

Field	Value
Original model	efficient_sam_s
Original source	GitHub
Original developer	Yunyang Xiong et al. (Meta AI Research)
Original model card	yformer/EfficientSAM
Original license	Apache-2.0

Model files

File	Description
efficient_sam_s_graviton_onnx_optimized.onnx	Arm-optimized INT8 model for deployment
example.py	Minimal inference example
config.yaml	Model I/O contract used by the example
metadata.yaml	Model metadata used by the example or Model Garden
pyproject.toml	Locked runtime environment definition (uv project)
.python-version	Python version pinned for the locked environment
uv.lock	Resolved dependency lockfile for reproducible runs
sample_input.jpg	Sample image used by the example
segmentation.json	Expected result of the example on sample_input.jpg
sample_output.png	Expected mask overlay of the example on sample_input.jpg

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	AWS Graviton G4
CPU / accelerator	Neoverse-V2 (arm64), CPU
OS	Linux — Ubuntu 24.04.4 LTS
Runtime	onnxruntime 1.26.0 (cpu)
Backend / delegate	CPU execution provider with MLAS + KleidiAI kernels (8 intra-op threads)
Batch size	1
Precision	INT8 weights (8-bit, per-channel, symmetric) and INT8 activations (8-bit, dynamic); image-encoder attention matmuls in dynamic INT8; mask-decoder attention and a few small mask-decoder linear layers in FP32
Runs	10 warmup + 100 measured

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
p50 latency	1656.103 ms	971.796 ms	1.70x
p90 latency	1658.775 ms	974.172 ms	1.70x
p99 latency	1663.513 ms	976.27 ms	1.70x
Throughput	0.6 FPS	1.03 FPS	1.72x
Model size	106.838 MB	31.818 MB	3.36x smaller
Peak memory	1321.57 MB	1556.53 MB	1.18x more

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	2500
Metric(s)	mIoU, Segm mAP@0.5:0.95, Segm mAP@0.5, Segm mAP@0.75
Evaluation runtime	ONNX Runtime

Accuracy results

Metric	Original / baseline	Arm-optimized	Change
mIoU	74.62%	74.71%	+0.09 pp
Segm mAP@0.5:0.95	42.81%	42.91%	+0.10 pp
Segm mAP@0.5	69.91%	70.17%	+0.26 pp
Segm mAP@0.75	45.03%	45.11%	+0.08 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-dynamic, symmetric, per-channel weights with dynamic INT8 activations; no calibration set required
Runtime/backend selection	Yes	ONNX Runtime CPU execution provider with MLAS and KleidiAI kernels
Graph/runtime compatibility updates	Yes	Activation quantize/dequantize chains removed so the INT8 linear kernels quantize their own inputs; image-encoder attention matmuls rewritten to dynamic INT8, with one scale per softmax row for attention × values
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

The environment was validated on Ubuntu 24.04, AArch64, with glibc 2.39 and no additional Apt packages.

This directory ships a uv-locked environment (pyproject.toml, .python-version, uv.lock). Install uv, then:

uv python install
uv sync --frozen

Run the example

uv run example.py

The ordinary run uses sample_input.jpg and the default box prompt 410 510 600 830. It prints a finite summary without changing repository files. To generate and compare the expected results:

uv run example.py --output-dir /tmp/efficient-sam-output
cmp /tmp/efficient-sam-output/sample_output.png sample_output.png
cmp /tmp/efficient-sam-output/segmentation.json segmentation.json

To run segmentation on another image, pass --image and provide --box in the resized 1024x1024 coordinate space. The box is specified as X1 Y1 X2 Y2:

uv run example.py \
  --image /path/to/image.jpg \
  --box 410 510 600 830 \
  --output-dir /tmp/efficient-sam-output

The example stretches the input image to 1024x1024; scale box coordinates from the original image before passing --box. sample_output.png contains the mask overlay and box prompt at the original input image resolution and aspect ratio; segmentation.json records the box and mask coverage.

Runtime notices

uv python install may warn that it cannot link a python3.14 executable into ~/.local/bin when one it does not manage is already there. uv still installs and uses its own interpreter, and the example produced byte-identical expected results. ONNX Runtime printed no warnings during validation.

Expected input

Property	Value
Input shape	[1, 3, 1024, 1024] image; [1, 1, 2, 2] point_coords; [1, 1, 2] point_labels
Input type	float32
Input range	[0.0, 1.0]
Preprocessing	Resize to 1024x1024 (bilinear, half-pixel centres, no antialiasing, no aspect-ratio preservation), convert uint8 [0-255] to float32 [0, 1]; no external normalization (applied inside the model). Box prompt is passed with --box X1 Y1 X2 Y2 as [[TL_x, TL_y], [BR_x, BR_y]] in 1024-pixel space with corner labels 2 (top-left) and 3 (bottom-right). The default prompt is 410 510 600 830, targeting the white armchair in sample_input.jpg.

Expected output

Property	Value
Output shape	[1, 1024, 1024]
Output type	Binary float mask — 1.0 inside the segmented region, 0.0 outside
Postprocessing	Threshold > 0.5 to produce the binary mask; the model already applies (logit > 0).float()
Saved files	sample_output.png: mask overlay at the original input image resolution; segmentation.json: mask coverage and the prompt used

Intended use

This model is intended for developers evaluating image-segmentation 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.
• The INT8 model uses more peak memory than FP32 and takes longer to load, so budget memory and startup time accordingly.
• 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

• Quantization recipe: PT2E dynamic INT8, symmetric per-channel weights. Attention probabilities over the image encoder's 4,096 keys get one uint8 scale per softmax row: with a single tensor-wide scale, near-uniform rows round to zero and the masks break. The mask decoder's attention and a few small decoder linear layers stay in FP32 to hold accuracy.
• The model uses ONNX opset 21 (blocked QuantizeLinear for the per-row scales); it was validated with ONNX Runtime 1.26.0.
• Requires Python 3.14.
• 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: efficient_sam_s by Yunyang Xiong et al. (Meta AI Research) - 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 Apache-2.0.

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.