BEVFormer: Optimized for Qualcomm Devices

Bevformer is a SOTA model of interest to the Auto BU.

This is based on the implementation of BEVFormer 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

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

Model Details

Model Type: Model_use_case.driver_assistance

Model Stats:

• Input preprocessing: Images must be ImageNet mean/std normalized before inference; the model does not normalize internally.
• Input resolution: 6 x 3 x 480 x 800
• Model checkpoint: bevformer_tiny_deformable_optimized_exp_86_epoch_24.pth
• Model size: 120 MB
• Number of parameters: 27M

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
BEVFormer	ONNX	float	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	885.831 ms	1 - 780 MB	NPU
BEVFormer	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	1342.611 ms	2 - 736 MB	NPU
BEVFormer	ONNX	float	Snapdragon® X2 Elite	895.635 ms	29 - 29 MB	NPU
BEVFormer	ONNX	float	Snapdragon® X Elite	1684.797 ms	50 - 50 MB	NPU
BEVFormer	ONNX	float	Snapdragon® 8 Gen 3 Mobile	1272.141 ms	30 - 1000 MB	NPU
BEVFormer	ONNX	float	Snapdragon® 8 Gen 1 Mobile	1865.385 ms	31 - 996 MB	NPU
BEVFormer	ONNX	float	Qualcomm® Dragonwing™ IQ-8275	1907.53 ms	29 - 62 MB	NPU
BEVFormer	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	1638.352 ms	0 - 53 MB	NPU
BEVFormer	ONNX	float	Qualcomm® QCS8450	1865.385 ms	31 - 996 MB	NPU
BEVFormer	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	1839.953 ms	29 - 61 MB	NPU
BEVFormer	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	1684.797 ms	50 - 50 MB	NPU
BEVFormer	ONNX	float	Qualcomm® Dragonwing™ Q-8750	1342.611 ms	2 - 736 MB	NPU
BEVFormer	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	876.881 ms	15 - 863 MB	NPU
BEVFormer	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	1203.975 ms	10 - 802 MB	NPU
BEVFormer	QNN_DLC	float	Snapdragon® X2 Elite	1020.819 ms	29 - 29 MB	NPU
BEVFormer	QNN_DLC	float	Snapdragon® X Elite	1712.873 ms	29 - 29 MB	NPU
BEVFormer	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	1337.472 ms	29 - 1003 MB	NPU
BEVFormer	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	1463.243 ms	2 - 988 MB	NPU
BEVFormer	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-8275	2024.129 ms	29 - 63 MB	NPU
BEVFormer	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	1618.435 ms	29 - 32 MB	NPU
BEVFormer	QNN_DLC	float	Qualcomm® SA8650P	2011.958 ms	26 - 863 MB	NPU
BEVFormer	QNN_DLC	float	Qualcomm® SA8255P	2011.958 ms	26 - 863 MB	NPU
BEVFormer	QNN_DLC	float	Qualcomm® QCS8450	1463.243 ms	2 - 988 MB	NPU
BEVFormer	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	1793.313 ms	29 - 62 MB	NPU
BEVFormer	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	1712.873 ms	29 - 29 MB	NPU
BEVFormer	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	1203.975 ms	10 - 802 MB	NPU
BEVFormer	QNN_DLC	float	Qualcomm® SA8295P	1764.945 ms	26 - 720 MB	NPU

License

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

References

• BEVFormer: Learning Bird's-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers
• 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.