CREStereo: Optimized for Qualcomm Devices

CREStereo (Cascaded Recurrent Network with Adaptive Correlation) is a CVPR 2022 Oral paper that achieves state-of-the-art stereo matching accuracy.

This is based on the implementation of CREStereo 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 2.50, ONNX Runtime 1.27.1	Download
QNN_DLC	float	Universal	QAIRT 2.50	Download
TFLITE	float	Universal	QAIRT 2.50	Download

For more device-specific assets and performance metrics, visit CREStereo 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 CREStereo on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.depth_estimation

Model Stats:

• Input: Rectified stereo pair — left and right RGB images
• Input resolution: 240x320
• Model checkpoint: CREStereo ETH3D pretrained (crestereo_eth3d.pt)
• Model size (float): 20.7 MB
• Number of parameters: 5.43M
• Output: Disparity map

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
CREStereo	ONNX	float	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	80.739 ms	2 - 809 MB	NPU
CREStereo	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	113.673 ms	2 - 1023 MB	NPU
CREStereo	ONNX	float	Snapdragon® X2 Elite	85.241 ms	2 - 2 MB	NPU
CREStereo	ONNX	float	Snapdragon® X Elite	200.83 ms	21 - 21 MB	NPU
CREStereo	ONNX	float	Snapdragon® 8 Gen 3 Mobile	150.423 ms	2 - 1327 MB	NPU
CREStereo	ONNX	float	Snapdragon® 8 Gen 1 Mobile	290.949 ms	3 - 1321 MB	NPU
CREStereo	ONNX	float	Qualcomm® Dragonwing™ IQ-8275	273.598 ms	2 - 6 MB	NPU
CREStereo	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	200.984 ms	0 - 26 MB	NPU
CREStereo	ONNX	float	Qualcomm® QCS8450	290.949 ms	3 - 1321 MB	NPU
CREStereo	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	268.392 ms	3 - 6 MB	NPU
CREStereo	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	200.83 ms	21 - 21 MB	NPU
CREStereo	ONNX	float	Qualcomm® Dragonwing™ Q-8750	113.673 ms	2 - 1023 MB	NPU
CREStereo	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	80.977 ms	0 - 720 MB	NPU
CREStereo	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	114.039 ms	1 - 930 MB	NPU
CREStereo	QNN_DLC	float	Snapdragon® X2 Elite	85.453 ms	2 - 2 MB	NPU
CREStereo	QNN_DLC	float	Snapdragon® X Elite	201.623 ms	2 - 2 MB	NPU
CREStereo	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	149.849 ms	2 - 1232 MB	NPU
CREStereo	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	291.569 ms	1 - 1219 MB	NPU
CREStereo	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-8275	274.15 ms	2 - 7 MB	NPU
CREStereo	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	201.719 ms	2 - 6 MB	NPU
CREStereo	QNN_DLC	float	Qualcomm® QCS8450	291.569 ms	1 - 1219 MB	NPU
CREStereo	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	268.345 ms	2 - 6 MB	NPU
CREStereo	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	201.623 ms	2 - 2 MB	NPU
CREStereo	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	114.039 ms	1 - 930 MB	NPU
CREStereo	QNN_DLC	float	Qualcomm® SA8295P	310.883 ms	3 - 974 MB	NPU
CREStereo	TFLITE	float	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	91.406 ms	0 - 1057 MB	NPU
CREStereo	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	124.803 ms	1 - 1251 MB	NPU
CREStereo	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	157.049 ms	1 - 1601 MB	NPU
CREStereo	TFLITE	float	Snapdragon® 8 Gen 1 Mobile	315.105 ms	1 - 1528 MB	NPU
CREStereo	TFLITE	float	Qualcomm® Dragonwing™ IQ-8275	283.148 ms	0 - 45 MB	NPU
CREStereo	TFLITE	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	218.402 ms	0 - 6 MB	NPU
CREStereo	TFLITE	float	Qualcomm® SA8775P	5476.969 ms	43 - 76 MB	CPU
CREStereo	TFLITE	float	Qualcomm® SA8650P	5476.969 ms	43 - 76 MB	CPU
CREStereo	TFLITE	float	Qualcomm® SA8255P	5476.969 ms	43 - 76 MB	CPU
CREStereo	TFLITE	float	Qualcomm® QCS8450	315.105 ms	1 - 1528 MB	NPU
CREStereo	TFLITE	float	Qualcomm® Dragonwing™ IQ-9075	280.439 ms	0 - 44 MB	NPU
CREStereo	TFLITE	float	Qualcomm® Dragonwing™ Q-8750	124.803 ms	1 - 1251 MB	NPU
CREStereo	TFLITE	float	Qualcomm® SA7255P	7168.524 ms	44 - 75 MB	CPU
CREStereo	TFLITE	float	Qualcomm® SA8295P	320.282 ms	1 - 1208 MB	NPU

License

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

References

• Practical Stereo Matching via Cascaded Recurrent Network with Adaptive Correlation
• 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.