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	ONNX Runtime 1.24.3	Download
QNN_DLC	float	Universal	QAIRT 2.45	Download
TFLITE	float	Universal		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:

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

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
CREStereo	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	3841.905 ms	77 - 97 MB	CPU
CREStereo	ONNX	float	Snapdragon® X2 Elite	2034.246 ms	179 - 179 MB	CPU
CREStereo	ONNX	float	Snapdragon® X Elite	6292.226 ms	184 - 184 MB	CPU
CREStereo	ONNX	float	Snapdragon® 8 Gen 3 Mobile	4583.867 ms	59 - 81 MB	CPU
CREStereo	ONNX	float	Qualcomm® QCS8550 (Proxy)	5846.538 ms	65 - 68 MB	CPU
CREStereo	ONNX	float	Qualcomm® QCS9075	3946.878 ms	95 - 97 MB	CPU
CREStereo	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	3982.126 ms	62 - 85 MB	CPU
CREStereo	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	2883.993 ms	90 - 179 MB	CPU
CREStereo	QNN_DLC	float	Snapdragon® X2 Elite	2727.641 ms	85 - 85 MB	CPU
CREStereo	QNN_DLC	float	Snapdragon® X Elite	9891.231 ms	85 - 85 MB	CPU
CREStereo	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	4974.327 ms	128 - 219 MB	CPU
CREStereo	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	10891.085 ms	126 - 210 MB	CPU
CREStereo	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	5571.348 ms	111 - 154 MB	CPU
CREStereo	QNN_DLC	float	Qualcomm® SA8775P	6625.184 ms	126 - 211 MB	CPU
CREStereo	QNN_DLC	float	Qualcomm® QCS9075	7734.589 ms	314 - 1217 MB	CPU
CREStereo	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	7865.986 ms	141 - 236 MB	CPU
CREStereo	QNN_DLC	float	Qualcomm® SA7255P	10891.085 ms	126 - 210 MB	CPU
CREStereo	QNN_DLC	float	Qualcomm® SA8295P	5456.534 ms	133 - 217 MB	CPU
CREStereo	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	3361.535 ms	389 - 480 MB	CPU
CREStereo	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	2917.477 ms	45 - 73 MB	CPU
CREStereo	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	4461.271 ms	46 - 76 MB	CPU
CREStereo	TFLITE	float	Qualcomm® QCS8275 (Proxy)	7591.053 ms	48 - 73 MB	CPU
CREStereo	TFLITE	float	Qualcomm® QCS8550 (Proxy)	5318.402 ms	47 - 60 MB	CPU
CREStereo	TFLITE	float	Qualcomm® SA8775P	6053.307 ms	48 - 73 MB	CPU
CREStereo	TFLITE	float	Qualcomm® QCS9075	6158.679 ms	47 - 153 MB	CPU
CREStereo	TFLITE	float	Qualcomm® QCS8450 (Proxy)	7052.932 ms	48 - 81 MB	CPU
CREStereo	TFLITE	float	Qualcomm® SA7255P	7591.053 ms	48 - 73 MB	CPU
CREStereo	TFLITE	float	Qualcomm® SA8295P	4082.577 ms	47 - 73 MB	CPU
CREStereo	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	3275.879 ms	47 - 75 MB	CPU

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.