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 2.45, ONNX Runtime 1.27.1	Download
TFLITE	float	Universal	QAIRT 2.45	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® X2 Elite	91.283 ms	2 - 2 MB	NPU
CREStereo	ONNX	float	Snapdragon® X Elite	211.043 ms	23 - 23 MB	NPU
CREStereo	ONNX	float	Snapdragon® 8 Gen 3 Mobile	147.106 ms	3 - 1522 MB	NPU
CREStereo	ONNX	float	Snapdragon® 8 Gen 1 Mobile	356.457 ms	4 - 1510 MB	NPU
CREStereo	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	206.458 ms	0 - 30 MB	NPU
CREStereo	ONNX	float	Qualcomm® QCS8450	356.457 ms	4 - 1510 MB	NPU
CREStereo	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	318.438 ms	2 - 5 MB	NPU
CREStereo	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	211.043 ms	23 - 23 MB	NPU
CREStereo	ONNX	float	Qualcomm® Dragonwing™ Q-8750	116.15 ms	3 - 1161 MB	NPU
CREStereo	ONNX	float	Snapdragon® 8 Elite Mobile	116.15 ms	3 - 1161 MB	NPU
CREStereo	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	88.4 ms	1 - 936 MB	NPU
CREStereo	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	152.812 ms	1 - 2171 MB	NPU
CREStereo	TFLITE	float	Snapdragon® 8 Gen 1 Mobile	378.156 ms	0 - 2318 MB	NPU
CREStereo	TFLITE	float	Qualcomm® Dragonwing™ QCS8275	813.393 ms	1 - 1598 MB	NPU
CREStereo	TFLITE	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	215.767 ms	0 - 14 MB	NPU
CREStereo	TFLITE	float	Qualcomm® SA8775P	281.954 ms	1 - 1596 MB	NPU
CREStereo	TFLITE	float	Qualcomm® SA8650P	281.954 ms	1 - 1596 MB	NPU
CREStereo	TFLITE	float	Qualcomm® SA8255P	281.954 ms	1 - 1596 MB	NPU
CREStereo	TFLITE	float	Qualcomm® QCS8450	378.156 ms	0 - 2318 MB	NPU
CREStereo	TFLITE	float	Qualcomm® Dragonwing™ IQ-9075	281.064 ms	0 - 44 MB	NPU
CREStereo	TFLITE	float	Qualcomm® Dragonwing™ Q-8750	123.347 ms	1 - 1653 MB	NPU
CREStereo	TFLITE	float	Qualcomm® SA7255P	813.393 ms	1 - 1598 MB	NPU
CREStereo	TFLITE	float	Qualcomm® SA8295P	357.7 ms	1 - 1858 MB	NPU
CREStereo	TFLITE	float	Snapdragon® 8 Elite Mobile	123.347 ms	1 - 1653 MB	NPU
CREStereo	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	88.415 ms	0 - 1271 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.