FCN-ResNet50: Optimized for Qualcomm Devices

FCN_ResNet50 is a machine learning model that can segment images from the COCO dataset. It uses ResNet50 as a backbone.

This is based on the implementation of FCN-ResNet50 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
ONNX	w8a8	Universal	QAIRT 2.37, ONNX Runtime 1.23.0	Download
QNN_DLC	float	Universal	QAIRT 2.42	Download
QNN_DLC	w8a8	Universal	QAIRT 2.42	Download
TFLITE	float	Universal	QAIRT 2.42, TFLite 2.17.0	Download
TFLITE	w8a8	Universal	QAIRT 2.42, TFLite 2.17.0	Download

For more device-specific assets and performance metrics, visit FCN-ResNet50 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 FCN-ResNet50 on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.semantic_segmentation

Model Stats:

• Model checkpoint: COCO_WITH_VOC_LABELS_V1
• Input resolution: 224x224
• Number of output classes: 21
• Number of parameters: 33.0M
• Model size (float): 126 MB
• Model size (w8a8): 32.2 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
FCN-ResNet50	ONNX	float	Snapdragon® X Elite	43.753 ms	63 - 63 MB	NPU
FCN-ResNet50	ONNX	float	Snapdragon® 8 Gen 3 Mobile	32.391 ms	2 - 314 MB	NPU
FCN-ResNet50	ONNX	float	Qualcomm® QCS8550 (Proxy)	43.223 ms	0 - 81 MB	NPU
FCN-ResNet50	ONNX	float	Qualcomm® QCS9075	74.573 ms	3 - 9 MB	NPU
FCN-ResNet50	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	26.542 ms	1 - 239 MB	NPU
FCN-ResNet50	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	20.443 ms	2 - 255 MB	NPU
FCN-ResNet50	ONNX	w8a8	Snapdragon® X Elite	13.859 ms	32 - 32 MB	NPU
FCN-ResNet50	ONNX	w8a8	Snapdragon® 8 Gen 3 Mobile	9.935 ms	0 - 231 MB	NPU
FCN-ResNet50	ONNX	w8a8	Qualcomm® QCS6490	921.792 ms	67 - 115 MB	CPU
FCN-ResNet50	ONNX	w8a8	Qualcomm® QCS8550 (Proxy)	13.896 ms	0 - 220 MB	NPU
FCN-ResNet50	ONNX	w8a8	Qualcomm® QCS9075	15.973 ms	1 - 4 MB	NPU
FCN-ResNet50	ONNX	w8a8	Qualcomm® QCM6690	824.792 ms	76 - 85 MB	CPU
FCN-ResNet50	ONNX	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	8.414 ms	1 - 156 MB	NPU
FCN-ResNet50	ONNX	w8a8	Snapdragon® 7 Gen 4 Mobile	700.625 ms	71 - 79 MB	CPU
FCN-ResNet50	ONNX	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	6.998 ms	1 - 197 MB	NPU
FCN-ResNet50	QNN_DLC	float	Snapdragon® X Elite	44.072 ms	3 - 3 MB	NPU
FCN-ResNet50	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	33.401 ms	0 - 382 MB	NPU
FCN-ResNet50	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	272.088 ms	1 - 305 MB	NPU
FCN-ResNet50	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	44.864 ms	3 - 6 MB	NPU
FCN-ResNet50	QNN_DLC	float	Qualcomm® SA8775P	72.053 ms	1 - 304 MB	NPU
FCN-ResNet50	QNN_DLC	float	Qualcomm® QCS9075	78.033 ms	3 - 8 MB	NPU
FCN-ResNet50	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	84.665 ms	1 - 276 MB	NPU
FCN-ResNet50	QNN_DLC	float	Qualcomm® SA7255P	272.088 ms	1 - 305 MB	NPU
FCN-ResNet50	QNN_DLC	float	Qualcomm® SA8295P	77.345 ms	0 - 217 MB	NPU
FCN-ResNet50	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	25.991 ms	3 - 323 MB	NPU
FCN-ResNet50	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	23.219 ms	3 - 330 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Snapdragon® X Elite	15.207 ms	1 - 1 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Snapdragon® 8 Gen 3 Mobile	11.408 ms	0 - 266 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Qualcomm® QCS6490	92.892 ms	1 - 3 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Qualcomm® QCS8275 (Proxy)	39.344 ms	1 - 205 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Qualcomm® QCS8550 (Proxy)	15.051 ms	1 - 3 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Qualcomm® SA8775P	14.56 ms	1 - 214 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Qualcomm® QCS9075	16.824 ms	3 - 5 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Qualcomm® QCM6690	407.762 ms	1 - 347 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Qualcomm® QCS8450 (Proxy)	23.958 ms	1 - 263 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Qualcomm® SA7255P	39.344 ms	1 - 205 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Qualcomm® SA8295P	21.86 ms	1 - 209 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	8.765 ms	1 - 198 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Snapdragon® 7 Gen 4 Mobile	25.963 ms	1 - 280 MB	NPU
FCN-ResNet50	QNN_DLC	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	7.52 ms	1 - 245 MB	NPU
FCN-ResNet50	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	33.035 ms	0 - 427 MB	NPU
FCN-ResNet50	TFLITE	float	Qualcomm® QCS8275 (Proxy)	272.13 ms	0 - 331 MB	NPU
FCN-ResNet50	TFLITE	float	Qualcomm® QCS8550 (Proxy)	44.874 ms	0 - 410 MB	NPU
FCN-ResNet50	TFLITE	float	Qualcomm® SA8775P	72.05 ms	0 - 333 MB	NPU
FCN-ResNet50	TFLITE	float	Qualcomm® QCS9075	77.817 ms	0 - 71 MB	NPU
FCN-ResNet50	TFLITE	float	Qualcomm® QCS8450 (Proxy)	84.565 ms	1 - 319 MB	NPU
FCN-ResNet50	TFLITE	float	Qualcomm® SA7255P	272.13 ms	0 - 331 MB	NPU
FCN-ResNet50	TFLITE	float	Qualcomm® SA8295P	77.444 ms	0 - 247 MB	NPU
FCN-ResNet50	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	27.043 ms	0 - 348 MB	NPU
FCN-ResNet50	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	20.946 ms	0 - 356 MB	NPU
FCN-ResNet50	TFLITE	w8a8	Snapdragon® 8 Gen 3 Mobile	10.969 ms	0 - 270 MB	NPU
FCN-ResNet50	TFLITE	w8a8	Qualcomm® QCS6490	94.9 ms	0 - 39 MB	NPU
FCN-ResNet50	TFLITE	w8a8	Qualcomm® QCS8275 (Proxy)	38.12 ms	0 - 204 MB	NPU
FCN-ResNet50	TFLITE	w8a8	Qualcomm® QCS8550 (Proxy)	14.445 ms	0 - 4 MB	NPU
FCN-ResNet50	TFLITE	w8a8	Qualcomm® SA8775P	14.511 ms	0 - 207 MB	NPU
FCN-ResNet50	TFLITE	w8a8	Qualcomm® QCS9075	15.341 ms	0 - 35 MB	NPU
FCN-ResNet50	TFLITE	w8a8	Qualcomm® QCM6690	433.506 ms	2 - 353 MB	NPU
FCN-ResNet50	TFLITE	w8a8	Qualcomm® QCS8450 (Proxy)	23.285 ms	0 - 269 MB	NPU
FCN-ResNet50	TFLITE	w8a8	Qualcomm® SA7255P	38.12 ms	0 - 204 MB	NPU
FCN-ResNet50	TFLITE	w8a8	Qualcomm® SA8295P	21.104 ms	0 - 209 MB	NPU
FCN-ResNet50	TFLITE	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	8.433 ms	0 - 200 MB	NPU
FCN-ResNet50	TFLITE	w8a8	Snapdragon® 7 Gen 4 Mobile	26.062 ms	0 - 281 MB	NPU
FCN-ResNet50	TFLITE	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	7.007 ms	0 - 244 MB	NPU

License

• The license for the original implementation of FCN-ResNet50 can be found here.

References

• Fully Convolutional Networks for 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.