3D-Deep-BOX: Optimized for Qualcomm Devices

3D Deep Box is a machine learning model that predicts 3D bounding boxes and classes of objects in an image.

This is based on the implementation of 3D-Deep-BOX 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 3D-Deep-BOX 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 3D-Deep-BOX on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.object_detection

Model Stats:

• Model checkpoint: YOLOv3-tiny
• Input resolution(YOLO): 224x640
• Input resolution(VGG): 224x224
• Number of parameters (Yolo2DDetection): 9.78M
• Model size (Yolo2DDetection) (float): 37.3 MB
• Number of parameters (VGG3DDetection): 46.1M
• Model size (VGG3DDetection) (float): 176 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
VGG3DDetection	ONNX	float	Snapdragon® X Elite	5.334 ms	87 - 87 MB	NPU
VGG3DDetection	ONNX	float	Snapdragon® 8 Gen 3 Mobile	3.972 ms	0 - 166 MB	NPU
VGG3DDetection	ONNX	float	Qualcomm® QCS8550 (Proxy)	5.271 ms	0 - 172 MB	NPU
VGG3DDetection	ONNX	float	Qualcomm® QCS9075	9.132 ms	1 - 5 MB	NPU
VGG3DDetection	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	3.227 ms	0 - 150 MB	NPU
VGG3DDetection	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	2.628 ms	0 - 151 MB	NPU
VGG3DDetection	QNN_DLC	float	Snapdragon® X Elite	5.24 ms	1 - 1 MB	NPU
VGG3DDetection	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	3.908 ms	0 - 217 MB	NPU
VGG3DDetection	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	33.44 ms	1 - 198 MB	NPU
VGG3DDetection	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	5.097 ms	1 - 2 MB	NPU
VGG3DDetection	QNN_DLC	float	Qualcomm® SA8775P	9.22 ms	1 - 198 MB	NPU
VGG3DDetection	QNN_DLC	float	Qualcomm® QCS9075	9.117 ms	1 - 3 MB	NPU
VGG3DDetection	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	10.039 ms	0 - 214 MB	NPU
VGG3DDetection	QNN_DLC	float	Qualcomm® SA7255P	33.44 ms	1 - 198 MB	NPU
VGG3DDetection	QNN_DLC	float	Qualcomm® SA8295P	10.075 ms	1 - 199 MB	NPU
VGG3DDetection	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	3.113 ms	1 - 202 MB	NPU
VGG3DDetection	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	2.477 ms	0 - 199 MB	NPU
VGG3DDetection	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	3.918 ms	0 - 263 MB	NPU
VGG3DDetection	TFLITE	float	Qualcomm® QCS8275 (Proxy)	33.446 ms	0 - 201 MB	NPU
VGG3DDetection	TFLITE	float	Qualcomm® QCS8550 (Proxy)	5.053 ms	0 - 3 MB	NPU
VGG3DDetection	TFLITE	float	Qualcomm® SA8775P	9.223 ms	0 - 202 MB	NPU
VGG3DDetection	TFLITE	float	Qualcomm® QCS9075	9.096 ms	2 - 93 MB	NPU
VGG3DDetection	TFLITE	float	Qualcomm® QCS8450 (Proxy)	10.059 ms	0 - 262 MB	NPU
VGG3DDetection	TFLITE	float	Qualcomm® SA7255P	33.446 ms	0 - 201 MB	NPU
VGG3DDetection	TFLITE	float	Qualcomm® SA8295P	10.105 ms	0 - 203 MB	NPU
VGG3DDetection	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	3.125 ms	0 - 206 MB	NPU
VGG3DDetection	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	2.505 ms	0 - 202 MB	NPU
Yolo2DDetection	ONNX	float	Snapdragon® X Elite	1.586 ms	23 - 23 MB	NPU
Yolo2DDetection	ONNX	float	Snapdragon® 8 Gen 3 Mobile	1.393 ms	0 - 103 MB	NPU
Yolo2DDetection	ONNX	float	Qualcomm® QCS8550 (Proxy)	1.65 ms	0 - 24 MB	NPU
Yolo2DDetection	ONNX	float	Qualcomm® QCS9075	2.452 ms	2 - 4 MB	NPU
Yolo2DDetection	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	1.058 ms	0 - 94 MB	NPU
Yolo2DDetection	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	0.938 ms	0 - 93 MB	NPU
Yolo2DDetection	QNN_DLC	float	Snapdragon® X Elite	1.484 ms	2 - 2 MB	NPU
Yolo2DDetection	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	1.058 ms	1 - 33 MB	NPU
Yolo2DDetection	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	7.567 ms	2 - 24 MB	NPU
Yolo2DDetection	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	1.357 ms	2 - 3 MB	NPU
Yolo2DDetection	QNN_DLC	float	Qualcomm® SA8775P	2.26 ms	1 - 26 MB	NPU
Yolo2DDetection	QNN_DLC	float	Qualcomm® QCS9075	2.179 ms	2 - 5 MB	NPU
Yolo2DDetection	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	3.868 ms	0 - 40 MB	NPU
Yolo2DDetection	QNN_DLC	float	Qualcomm® SA7255P	7.567 ms	2 - 24 MB	NPU
Yolo2DDetection	QNN_DLC	float	Qualcomm® SA8295P	3.575 ms	0 - 23 MB	NPU
Yolo2DDetection	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	0.879 ms	2 - 23 MB	NPU
Yolo2DDetection	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	0.712 ms	2 - 26 MB	NPU
Yolo2DDetection	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	1.038 ms	0 - 67 MB	NPU
Yolo2DDetection	TFLITE	float	Qualcomm® QCS8275 (Proxy)	7.553 ms	0 - 25 MB	NPU
Yolo2DDetection	TFLITE	float	Qualcomm® QCS8550 (Proxy)	1.284 ms	0 - 2 MB	NPU
Yolo2DDetection	TFLITE	float	Qualcomm® SA8775P	2.223 ms	0 - 28 MB	NPU
Yolo2DDetection	TFLITE	float	Qualcomm® QCS9075	2.108 ms	0 - 27 MB	NPU
Yolo2DDetection	TFLITE	float	Qualcomm® QCS8450 (Proxy)	3.849 ms	0 - 67 MB	NPU
Yolo2DDetection	TFLITE	float	Qualcomm® SA7255P	7.553 ms	0 - 25 MB	NPU
Yolo2DDetection	TFLITE	float	Qualcomm® SA8295P	3.541 ms	0 - 26 MB	NPU
Yolo2DDetection	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	0.895 ms	0 - 30 MB	NPU
Yolo2DDetection	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	0.689 ms	0 - 28 MB	NPU

License

• The license for the original implementation of 3D-Deep-BOX can be found here.

References

• 3D Bounding Box Estimation Using Deep Learning and Geometry
• 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.