YoloV5-Face: Optimized for Qualcomm Devices

YoloV5-Face-Nano is a lightweight single-class face detector based on the YoloV5 backbone (StemBlock + ShuffleV2Block). Each detected face is accompanied by 5 facial landmark predictions (left eye, right eye, nose, left mouth corner, right mouth corner). Trained on the WiderFace dataset.

This is based on the implementation of YoloV5-Face 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

Due to licensing restrictions, we cannot distribute pre-exported model assets for this model. 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

See our repository for YoloV5-Face on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.object_detection

Model Stats:

• Input resolution: 640x640
• Model checkpoint: yolov5n-face.pt
• Model size (float): 7.21 MB
• Model size (w8a16): 3.68 MB
• Number of parameters: 1.87M

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
YoloV5-Face	ONNX	float	Snapdragon® X2 Elite	6.203 ms	9 - 9 MB	NPU
YoloV5-Face	ONNX	float	Snapdragon® X Elite	13.752 ms	9 - 9 MB	NPU
YoloV5-Face	ONNX	float	Snapdragon® 8 Gen 3 Mobile	7.522 ms	10 - 223 MB	NPU
YoloV5-Face	ONNX	float	Snapdragon® 8 Gen 1 Mobile	26.283 ms	8 - 242 MB	NPU
YoloV5-Face	ONNX	float	Qualcomm® Dragonwing™ IQ-8275	11.12 ms	6 - 9 MB	NPU
YoloV5-Face	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	13.074 ms	6 - 124 MB	NPU
YoloV5-Face	ONNX	float	Qualcomm® QCS8450	26.283 ms	8 - 242 MB	NPU
YoloV5-Face	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	14.41 ms	6 - 8 MB	NPU
YoloV5-Face	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	13.752 ms	9 - 9 MB	NPU
YoloV5-Face	ONNX	float	Qualcomm® Dragonwing™ Q-8750	6.461 ms	3 - 193 MB	NPU
YoloV5-Face	ONNX	float	Snapdragon® 8 Elite Mobile	6.461 ms	3 - 193 MB	NPU
YoloV5-Face	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	5.95 ms	3 - 197 MB	NPU
YoloV5-Face	ONNX	w8a16	Snapdragon® X2 Elite	6.189 ms	4 - 4 MB	NPU
YoloV5-Face	ONNX	w8a16	Snapdragon® X Elite	14.571 ms	4 - 4 MB	NPU
YoloV5-Face	ONNX	w8a16	Snapdragon® 8 Gen 3 Mobile	7.979 ms	4 - 271 MB	NPU
YoloV5-Face	ONNX	w8a16	Snapdragon® 8 Gen 1 Mobile	16.502 ms	4 - 267 MB	NPU
YoloV5-Face	ONNX	w8a16	Qualcomm® Dragonwing™ QCS6490	35.893 ms	3 - 6 MB	NPU
YoloV5-Face	ONNX	w8a16	Qualcomm® Dragonwing™ IQ-8275	10.82 ms	3 - 7 MB	NPU
YoloV5-Face	ONNX	w8a16	Qualcomm® Dragonwing™ QCS8550 (Proxy)	13.924 ms	3 - 137 MB	NPU
YoloV5-Face	ONNX	w8a16	Qualcomm® QCS8450	16.502 ms	4 - 267 MB	NPU
YoloV5-Face	ONNX	w8a16	Qualcomm® Dragonwing™ IQ-9075	15.075 ms	3 - 6 MB	NPU
YoloV5-Face	ONNX	w8a16	Qualcomm® Dragonwing™ IQ-X7181	14.571 ms	4 - 4 MB	NPU
YoloV5-Face	ONNX	w8a16	Qualcomm® Dragonwing™ Q-6690	50.308 ms	4 - 275 MB	NPU
YoloV5-Face	ONNX	w8a16	Qualcomm® Dragonwing™ Q-7790	16.303 ms	4 - 277 MB	NPU
YoloV5-Face	ONNX	w8a16	Qualcomm® Dragonwing™ Q-8750	6.595 ms	2 - 235 MB	NPU
YoloV5-Face	ONNX	w8a16	Snapdragon® 8 Elite Mobile	6.595 ms	2 - 235 MB	NPU
YoloV5-Face	ONNX	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	6.024 ms	1 - 240 MB	NPU
YoloV5-Face	ONNX	w8a16	Snapdragon® 7 Gen 4 Mobile	16.303 ms	4 - 277 MB	NPU
YoloV5-Face	QNN_DLC	float	Snapdragon® X2 Elite	2.372 ms	5 - 5 MB	NPU
YoloV5-Face	QNN_DLC	float	Snapdragon® X Elite	4.654 ms	5 - 5 MB	NPU
YoloV5-Face	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	2.88 ms	5 - 91 MB	NPU
YoloV5-Face	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	8.303 ms	2 - 103 MB	NPU
YoloV5-Face	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-8275	5.027 ms	5 - 13 MB	NPU
YoloV5-Face	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	4.048 ms	4 - 7 MB	NPU
YoloV5-Face	QNN_DLC	float	Qualcomm® SA8775P	5.473 ms	0 - 65 MB	NPU
YoloV5-Face	QNN_DLC	float	Qualcomm® SA8650P	5.473 ms	0 - 65 MB	NPU
YoloV5-Face	QNN_DLC	float	Qualcomm® SA8255P	5.473 ms	0 - 65 MB	NPU
YoloV5-Face	QNN_DLC	float	Qualcomm® QCS8450	8.303 ms	2 - 103 MB	NPU
YoloV5-Face	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	5.179 ms	5 - 12 MB	NPU
YoloV5-Face	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	4.654 ms	5 - 5 MB	NPU
YoloV5-Face	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	2.203 ms	5 - 72 MB	NPU
YoloV5-Face	QNN_DLC	float	Qualcomm® SA7255P	12.807 ms	0 - 62 MB	NPU
YoloV5-Face	QNN_DLC	float	Qualcomm® SA8295P	9.049 ms	0 - 72 MB	NPU
YoloV5-Face	QNN_DLC	float	Snapdragon® 8 Elite Mobile	2.203 ms	5 - 72 MB	NPU
YoloV5-Face	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	1.796 ms	5 - 72 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Snapdragon® X2 Elite	2.094 ms	2 - 2 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Snapdragon® X Elite	5.046 ms	2 - 2 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Snapdragon® 8 Gen 3 Mobile	3.039 ms	2 - 106 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Snapdragon® 8 Gen 1 Mobile	4.941 ms	2 - 104 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® Dragonwing™ QCS6490	14.876 ms	2 - 7 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® Dragonwing™ IQ-8275	4.384 ms	2 - 8 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® Dragonwing™ QCS8550 (Proxy)	4.455 ms	2 - 4 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® SA8775P	5.121 ms	1 - 79 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® SA8650P	5.121 ms	1 - 79 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® SA8255P	5.121 ms	1 - 79 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® QCS8450	4.941 ms	2 - 104 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® Dragonwing™ IQ-9075	4.908 ms	2 - 7 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® Dragonwing™ IQ-X7181	5.046 ms	2 - 2 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-6690	20.511 ms	2 - 198 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-7790	4.933 ms	2 - 201 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-8750	2.079 ms	2 - 203 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® SA7255P	8.446 ms	1 - 80 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Qualcomm® SA8295P	5.76 ms	0 - 76 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Snapdragon® 8 Elite Mobile	2.079 ms	2 - 203 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	1.677 ms	2 - 204 MB	NPU
YoloV5-Face	QNN_DLC	w8a16	Snapdragon® 7 Gen 4 Mobile	4.933 ms	2 - 201 MB	NPU
YoloV5-Face	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	2.912 ms	0 - 91 MB	NPU
YoloV5-Face	TFLITE	float	Snapdragon® 8 Gen 1 Mobile	8.144 ms	2 - 107 MB	NPU
YoloV5-Face	TFLITE	float	Qualcomm® Dragonwing™ IQ-8275	5.046 ms	1 - 15 MB	NPU
YoloV5-Face	TFLITE	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	3.979 ms	0 - 3 MB	NPU
YoloV5-Face	TFLITE	float	Qualcomm® SA8775P	5.449 ms	1 - 76 MB	NPU
YoloV5-Face	TFLITE	float	Qualcomm® SA8650P	5.449 ms	1 - 76 MB	NPU
YoloV5-Face	TFLITE	float	Qualcomm® SA8255P	5.449 ms	1 - 76 MB	NPU
YoloV5-Face	TFLITE	float	Qualcomm® QCS8450	8.144 ms	2 - 107 MB	NPU
YoloV5-Face	TFLITE	float	Qualcomm® Dragonwing™ IQ-9075	5.171 ms	0 - 13 MB	NPU
YoloV5-Face	TFLITE	float	Qualcomm® Dragonwing™ Q-8750	2.208 ms	0 - 69 MB	NPU
YoloV5-Face	TFLITE	float	Qualcomm® SA7255P	12.765 ms	1 - 68 MB	NPU
YoloV5-Face	TFLITE	float	Qualcomm® SA8295P	9.045 ms	2 - 77 MB	NPU
YoloV5-Face	TFLITE	float	Snapdragon® 8 Elite Mobile	2.208 ms	0 - 69 MB	NPU
YoloV5-Face	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	1.693 ms	0 - 73 MB	NPU

License

• The license for the original implementation of YoloV5-Face can be found here.

References

• YOLO5Face: Why Reinventing a Face Detector
• 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.