Arm/depth-anything-v2-small-int8-xnnpack-executorch-vivo-x300

Depth Anything V2 Small optimized for Arm-based mobile CPUs with SME2

An INT8-quantized version of Depth Anything V2 Small for monocular depth estimation, exported to ExecuTorch (.pte) and optimized for Arm-based mobile CPUs with SME2.

Summary

This repository contains an Arm-optimized version of Depth Anything V2 Small for monocular depth estimation, quantized to INT8 via static post-training quantization (PTQ) with per-channel symmetric weights and per-tensor symmetric activations. The model is provided in ExecuTorch (.pte) format, targeting Mobile CPU systems.

This version is intended to demonstrate efficient inference on Arm-based platforms while preserving the original model's intended behavior. Arm evaluated the model and measured the published performance and memory results on a representative evaluation target. The included Python example and locked uv environment target a Linux arm64 host; the representative evaluation target measurements must not be interpreted as Linux host performance results.

Key results

Area	Result
Model format	ExecuTorch (.pte)
Target device class	Mobile CPU
Reference device	vivo X300 (aarch64, 8-core CPU, Android 16 / OriginOS 6)
Example environment	AWS Graviton, Ubuntu arm64
Primary performance result	p50 latency 727.726 ms, 1.37 FPS (1.84x faster than baseline)
Accuracy result	RMSE 0.3474 m, AbsRel 0.0600, delta1 (< 1.25) 96.35%
Size / memory result	26.564 MB (3.56x smaller), peak memory 380.10 MB (1.16x less)

Original model

Field	Value
Original model	Depth Anything V2 Small
Original source	Hugging Face
Original developer	Depth Anything team (HKU / TikTok)
Original model card	depth-anything/Depth-Anything-V2-Small-hf
Original license	Apache-2.0

Model files

File	Description
depth_anything_v2_small_executorch_optimized.pte	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 input/output contract used by the example
benchmarks/	FP32 baseline and Arm-optimized benchmark records

Performance

Performance was measured on the vivo X300 reference configuration below. The Graviton environment documented for example.py verifies portability of the inference flow but does not provide or imply equivalent Graviton performance.

Reference configuration

Field	Value
Device / platform	vivo X300
CPU architecture	aarch64
CPU	C1-Ultra, C1-Premium, C1-Pro, 8 cores
System memory	16 GB
OS	Android 16 / OriginOS 6
Runtime	ExecuTorch 1.1.0
Execution backend	CPU (XNNPACK, KleidiAI)
Precision	INT8, static PTQ — per-channel symmetric weights, per-tensor symmetric activations (zero-point 0)
Batch size	1
Threads	1
Input resolution	518x686
Runs / warmup	50 / 10

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
Model size (MB)	94.516	26.564	3.56x smaller
End-to-end latency p50 (ms)	1336.611	727.726	1.84x faster
End-to-end latency p90 (ms)	1344.712	735.644	1.83x faster
End-to-end latency p99 (ms)	1350.503	737.494	1.83x faster
Model load time (ms)	30.270	30.547	Slower
Time to first inference (ms)	1122.838	750.388	1.50x faster
Peak memory, USS (MB)	439.43	380.10	1.16x less
Frames per second	0.80	1.37	1.71x

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	NYU Depth V2
Split	Test (official raw-depth test split)
Sample count	654
Metric(s)	RMSE (m), AbsRel, delta1 (< 1.25)
Prediction alignment	Per-image affine scale-and-shift fit to ground truth in disparity space before computing metrics
Runtime	ExecuTorch 1.1.0 (Arm-optimized) / PyTorch (baseline), CPU with XNNPACK and KleidiAI

Accuracy results

Metric	Original / baseline	Arm-optimized	Change
AbsRel (lower is better)	0.0512	0.0600	+0.0088
RMSE, m (lower is better)	0.3108	0.3474	+0.0366
delta1 (< 1.25), % (higher is better)	97.32	96.35	-0.97 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	Loaded from Hugging Face with eager attention, wrapped to return a plain depth tensor, and converted to ExecuTorch .pte via the shared PT2E export pipeline
Quantization	Yes	INT8 static PTQ with histogram observers — per-channel symmetric weights, per-tensor symmetric activations; calibrated on 100 NYU Depth V2 samples
Runtime/backend selection	Yes	XNNPACK + KleidiAI delegate on the ExecuTorch CPU backend
Graph/runtime compatibility updates	Yes	Kept aten.linear undecomposed for direct XNNPACK delegation and used group-based partitioning; no post-quantization graph surgery was required
Accuracy validation	Yes	Compared against the original model under the same evaluation protocol
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

Note: The Python/uv example runs on AWS Graviton (Ubuntu arm64) to confirm runtime compatibility only, and is intended as a guideline for building an equivalent run script on Mobile CPU systems.

The reproducible Python example environment targets AWS Graviton running Ubuntu on arm64. Dependencies are declared in pyproject.toml; its lock includes Linux aarch64 wheels for the ExecuTorch/XNNPACK runtime and preprocessing packages. Resolve and install them with uv:

uv python install
uv sync --frozen

Run the example

uv run example.py

The ordinary command prints a finite disparity summary and does not overwrite the committed expected results. To write new output files, pass an explicit directory:

uv run example.py --output-dir outputs

The script reads sample_input.jpg, runs inference with depth_anything_v2_small_executorch_optimized.pte, and optionally writes:

• sample_output.png — grayscale disparity map at the original image resolution (255 = nearest)
• depth.json — input/output metadata and raw disparity statistics

Expected input

Property	Value
Shape	[1, 3, 518, 686]
Dtype	float32
Range	[0.0, 1.0] before normalization; the tensor passed to the model is normalized (approximately [-2.12, 2.64])
Preprocessing	Convert RGB uint8 HWC pixels to NumPy float64 in [0, 1] by dividing by 255.0; resize directly to 518x686 (H x W) with OpenCV INTER_CUBIC; normalize in NumPy with ImageNet mean [0.485, 0.456, 0.406] and std [0.229, 0.224, 0.225]; transpose to CHW, cast to float32, and add the batch dimension. This matches the calibration and evaluation pipeline.

Expected output

Property	Value
Shape	[1, 518, 686]
Format	Relative disparity map (higher values are closer), not metric depth
Postprocessing	squeeze the singleton channel dimension if present; apply per-sample min-max normalization to [0, 1]; bilinearly resize to the original image resolution for visualization and save as an 8-bit grayscale image

Intended use

This model is intended for developers evaluating monocular depth estimation workloads on Arm-based platforms. It is suitable as a reference implementation for benchmarking, prototyping, and integration exploration.

Limitations

• Evaluated on the 654-image official NYU Depth V2 test split, which contains indoor scenes; outdoor and highly out-of-distribution domains are not covered by these results.
• The input size is fixed at 518x686. This artifact matches the paper-aligned NYU 4:3 geometry; other aspect ratios require a separately exported static shape.
• The output is relative disparity, not metric depth. Converting it to metric depth requires affine scale-and-shift alignment outside the model.
• Performance was measured on a vivo X300; other Arm devices may run slower or faster and may have different XNNPACK or KleidiAI delegation coverage.
• The Python/uv example targets AWS Graviton on Ubuntu arm64. It establishes runtime compatibility only; no Graviton latency or memory benchmark is reported here.

Additional notes

Quantization was performed with PT2E static quantization using histogram observers, INT8 per-channel symmetric weights, and per-tensor symmetric activations with zero-point 0. No layers were kept in FP32; the whole model was quantized. Calibration used 100 samples from a Hugging Face-compatible NYU Depth V2 subset, separate from the 654-image official test split used for evaluation.

This artifact uses a 518x686 (H x W) input, matching the paper-aligned NYU 4:3 geometry, rather than a square 518x518 shape. The square shape requires aspect-distorting resizing, while 518x686 processes more pixels and retains the intended geometry.

• 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: Depth Anything V2 Small by Depth Anything team (HKU / TikTok) - 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.