StereoNet: Optimized for Qualcomm Devices

StereoNet is an end-to-end deep architecture for real-time stereo matching that produces high-quality, edge-preserved disparity maps from a rectified stereo image pair.

This is based on the implementation of StereoNet found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.

Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.

Getting Started

There are two ways to deploy this model on your device:

Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

Runtime	Precision	Chipset	SDK Versions	Download
ONNX	float	Universal	QAIRT 2.42, ONNX Runtime 1.24.1	Download
QNN_DLC	float	Universal	QAIRT 2.43	Download
TFLITE	float	Universal	QAIRT 2.43, TFLite 2.17.0	Download

For more device-specific assets and performance metrics, visit StereoNet on Qualcomm® AI Hub.

Option 2: Export with Custom Configurations

Use the Qualcomm® AI Hub Models Python library to compile and export the model with your own:

• Custom weights (e.g., fine-tuned checkpoints)
• Custom input shapes
• Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for StereoNet on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.depth_estimation

Model Stats:

• Model checkpoint: KeystoneDepth (epoch=21-step=696366.ckpt)
• Input resolution: 786x490
• Number of parameters: 1.94M
• Model size (float): 7.41 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
StereoNet	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	184.46 ms	6 - 1360 MB	NPU
StereoNet	ONNX	float	Snapdragon® X2 Elite	180.253 ms	20 - 20 MB	NPU
StereoNet	ONNX	float	Snapdragon® X Elite	329.003 ms	20 - 20 MB	NPU
StereoNet	ONNX	float	Snapdragon® 8 Gen 3 Mobile	261.796 ms	6 - 1980 MB	NPU
StereoNet	ONNX	float	Qualcomm® QCS8550 (Proxy)	391.508 ms	0 - 1078 MB	NPU
StereoNet	ONNX	float	Qualcomm® QCS9075	514.272 ms	3 - 6 MB	NPU
StereoNet	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	219.257 ms	3 - 1326 MB	NPU
StereoNet	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	164.683 ms	3 - 1355 MB	NPU
StereoNet	QNN_DLC	float	Snapdragon® X2 Elite	163.103 ms	3 - 3 MB	NPU
StereoNet	QNN_DLC	float	Snapdragon® X Elite	313.701 ms	3 - 3 MB	NPU
StereoNet	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	246.971 ms	3 - 1992 MB	NPU
StereoNet	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	1207.886 ms	2 - 1362 MB	NPU
StereoNet	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	336.051 ms	3 - 5 MB	NPU
StereoNet	QNN_DLC	float	Qualcomm® SA8775P	1789.115 ms	1 - 1362 MB	NPU
StereoNet	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	766.006 ms	3 - 2149 MB	NPU
StereoNet	QNN_DLC	float	Qualcomm® SA7255P	1207.886 ms	2 - 1362 MB	NPU
StereoNet	QNN_DLC	float	Qualcomm® SA8295P	474.277 ms	0 - 1499 MB	NPU
StereoNet	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	202.351 ms	0 - 1333 MB	NPU
StereoNet	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	230.663 ms	72 - 1694 MB	NPU
StereoNet	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	312.911 ms	73 - 2203 MB	NPU
StereoNet	TFLITE	float	Qualcomm® QCS8275 (Proxy)	1322.775 ms	73 - 1627 MB	NPU
StereoNet	TFLITE	float	Qualcomm® QCS8550 (Proxy)	456.107 ms	74 - 78 MB	NPU
StereoNet	TFLITE	float	Qualcomm® SA8775P	526.555 ms	74 - 1628 MB	NPU
StereoNet	TFLITE	float	Qualcomm® QCS8450 (Proxy)	766.181 ms	74 - 2482 MB	NPU
StereoNet	TFLITE	float	Qualcomm® SA7255P	1322.775 ms	73 - 1627 MB	NPU
StereoNet	TFLITE	float	Qualcomm® SA8295P	562.674 ms	74 - 1726 MB	NPU
StereoNet	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	254.335 ms	72 - 1657 MB	NPU

License

• The license for the original implementation of StereoNet can be found here.

References

• StereoNet: Guided Hierarchical Refinement for Real-Time Edge-Aware Depth Prediction
• Source Model Implementation

Community

• Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
• For questions or feedback please reach out to us.