CVT: Optimized for Qualcomm Devices

Cross-View Transformer generates real-time bird's-eye view maps from multiple vehicle cameras for autonomous driving.

This is based on the implementation of CVT 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 2.37, ONNX Runtime 1.23.0	Download
QNN_DLC	float	Universal	QAIRT 2.42	Download
TFLITE	float	Universal	QAIRT 2.42, TFLite 2.17.0	Download

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

Model Details

Model Type: Model_use_case.driver_assistance

Model Stats:

• Model checkpoint: vehicles_50k.pt
• Inference latency: RealTime
• Input resolution: 1x6x3x224x480
• Number of parameters: 1.33M
• Model size (float): 5.18 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
CVT	ONNX	float	Snapdragon® X Elite	337.851 ms	20 - 20 MB	NPU
CVT	ONNX	float	Snapdragon® 8 Gen 3 Mobile	307.125 ms	8 - 2453 MB	NPU
CVT	ONNX	float	Qualcomm® QCS8550 (Proxy)	407.302 ms	0 - 24 MB	NPU
CVT	ONNX	float	Qualcomm® QCS9075	331.746 ms	7 - 10 MB	NPU
CVT	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	259.644 ms	4 - 2088 MB	NPU
CVT	QNN_DLC	float	Snapdragon® X Elite	281.503 ms	7 - 7 MB	NPU
CVT	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	221.782 ms	8 - 2556 MB	NPU
CVT	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	423.067 ms	0 - 2186 MB	NPU
CVT	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	329.044 ms	8 - 578 MB	NPU
CVT	QNN_DLC	float	Qualcomm® SA8775P	292.094 ms	0 - 2186 MB	NPU
CVT	QNN_DLC	float	Qualcomm® QCS9075	323.368 ms	7 - 17 MB	NPU
CVT	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	439.183 ms	7 - 2650 MB	NPU
CVT	QNN_DLC	float	Qualcomm® SA7255P	423.067 ms	0 - 2186 MB	NPU
CVT	QNN_DLC	float	Qualcomm® SA8295P	332.938 ms	0 - 2240 MB	NPU
CVT	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	184.599 ms	7 - 2173 MB	NPU
CVT	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	169.256 ms	7 - 2259 MB	NPU
CVT	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	224.811 ms	0 - 2562 MB	NPU
CVT	TFLITE	float	Qualcomm® QCS8275 (Proxy)	426.207 ms	0 - 2193 MB	NPU
CVT	TFLITE	float	Qualcomm® QCS8550 (Proxy)	339.717 ms	0 - 4 MB	NPU
CVT	TFLITE	float	Qualcomm® SA8775P	297.388 ms	0 - 2193 MB	NPU
CVT	TFLITE	float	Qualcomm® QCS9075	325.057 ms	0 - 35 MB	NPU
CVT	TFLITE	float	Qualcomm® QCS8450 (Proxy)	451.485 ms	0 - 2731 MB	NPU
CVT	TFLITE	float	Qualcomm® SA7255P	426.207 ms	0 - 2193 MB	NPU
CVT	TFLITE	float	Qualcomm® SA8295P	342.185 ms	0 - 2289 MB	NPU
CVT	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	188.96 ms	0 - 2243 MB	NPU
CVT	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	172.317 ms	0 - 2287 MB	NPU

License

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

References

• Cross-view Transformers for real-time Map-view Semantic Segmentation
• 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.