MaskRCNN: Optimized for Qualcomm Devices

Mask R-CNN is a machine learning model that extends Faster R-CNN to perform instance segmentation by detecting objects in an image while simultaneously generating a high-quality segmentation mask for each instance. It adds a branch for predicting segmentation masks in parallel with the existing branch for bounding box recognition.

This is based on the implementation of MaskRCNN 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 1.30.0	Download
QNN_DLC	float	Universal	QAIRT 2.50	Download

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

Model Details

Model Type: Model_use_case.semantic_segmentation

Model Stats:

• Input resolution: 800x800
• Model checkpoint: Mask R-CNN ResNet-50 FPN V2
• Model size (float): 177 MB
• Number of output classes: 91
• Number of parameters: 46.4M

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
proposal_generator	ONNX	float	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	41.885 ms	124 - 1350 MB	NPU
proposal_generator	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	52.793 ms	127 - 1502 MB	NPU
proposal_generator	ONNX	float	Snapdragon® X2 Elite	35.289 ms	272 - 272 MB	NPU
proposal_generator	ONNX	float	Snapdragon® 8 Gen 3 Mobile	53.215 ms	140 - 2776 MB	NPU
proposal_generator	ONNX	float	Snapdragon® 8 Gen 1 Mobile	106.773 ms	129 - 2697 MB	NPU
proposal_generator	ONNX	float	Qualcomm® Dragonwing™ IQ-8275	109.484 ms	43 - 54 MB	NPU
proposal_generator	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	74.216 ms	122 - 198 MB	NPU
proposal_generator	ONNX	float	Qualcomm® QCS8450	106.773 ms	129 - 2697 MB	NPU
proposal_generator	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	102.802 ms	7 - 18 MB	NPU
proposal_generator	ONNX	float	Qualcomm® Dragonwing™ Q-8750	52.793 ms	127 - 1502 MB	NPU
proposal_generator	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	41.682 ms	7 - 1444 MB	NPU
proposal_generator	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	54.022 ms	0 - 1619 MB	NPU
proposal_generator	QNN_DLC	float	Snapdragon® X2 Elite	41.472 ms	7 - 7 MB	NPU
proposal_generator	QNN_DLC	float	Snapdragon® X Elite	87.099 ms	7 - 7 MB	NPU
proposal_generator	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	56.941 ms	7 - 2158 MB	NPU
proposal_generator	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	107.846 ms	2 - 2591 MB	NPU
proposal_generator	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-8275	116.65 ms	7 - 73 MB	NPU
proposal_generator	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	88.522 ms	7 - 75 MB	NPU
proposal_generator	QNN_DLC	float	Qualcomm® SA8650P	122.238 ms	1 - 1758 MB	NPU
proposal_generator	QNN_DLC	float	Qualcomm® SA8255P	122.238 ms	1 - 1758 MB	NPU
proposal_generator	QNN_DLC	float	Qualcomm® QCS8450	107.846 ms	2 - 2591 MB	NPU
proposal_generator	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	113.631 ms	7 - 72 MB	NPU
proposal_generator	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	87.099 ms	7 - 7 MB	NPU
proposal_generator	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	54.022 ms	0 - 1619 MB	NPU
proposal_generator	QNN_DLC	float	Qualcomm® SA8295P	114.365 ms	0 - 1300 MB	NPU
roi_head	ONNX	float	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	48.329 ms	223 - 588 MB	NPU
roi_head	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	62.73 ms	227 - 591 MB	NPU
roi_head	ONNX	float	Snapdragon® X2 Elite	44.381 ms	310 - 310 MB	NPU
roi_head	ONNX	float	Snapdragon® 8 Gen 3 Mobile	70.548 ms	196 - 643 MB	NPU
roi_head	ONNX	float	Snapdragon® 8 Gen 1 Mobile	144.155 ms	234 - 652 MB	NPU
roi_head	ONNX	float	Qualcomm® Dragonwing™ IQ-8275	154.974 ms	108 - 208 MB	NPU
roi_head	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	87.35 ms	227 - 256 MB	NPU
roi_head	ONNX	float	Qualcomm® QCS8450	144.155 ms	234 - 652 MB	NPU
roi_head	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	116.225 ms	111 - 210 MB	NPU
roi_head	ONNX	float	Qualcomm® Dragonwing™ Q-8750	62.73 ms	227 - 591 MB	NPU
roi_head	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	85.72 ms	11 - 558 MB	NPU
roi_head	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	111.168 ms	8 - 543 MB	NPU
roi_head	QNN_DLC	float	Snapdragon® X2 Elite	91.463 ms	52 - 52 MB	NPU
roi_head	QNN_DLC	float	Snapdragon® X Elite	240.1 ms	52 - 52 MB	NPU
roi_head	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	164.622 ms	48 - 681 MB	NPU
roi_head	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	306.495 ms	11 - 610 MB	NPU
roi_head	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-8275	241.948 ms	52 - 107 MB	NPU
roi_head	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	239.408 ms	52 - 612 MB	NPU
roi_head	QNN_DLC	float	Qualcomm® SA8650P	261.792 ms	49 - 777 MB	NPU
roi_head	QNN_DLC	float	Qualcomm® SA8255P	261.792 ms	49 - 777 MB	NPU
roi_head	QNN_DLC	float	Qualcomm® QCS8450	306.495 ms	11 - 610 MB	NPU
roi_head	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	260.331 ms	52 - 107 MB	NPU
roi_head	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	240.1 ms	52 - 52 MB	NPU
roi_head	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	111.168 ms	8 - 543 MB	NPU
roi_head	QNN_DLC	float	Qualcomm® SA8295P	289.176 ms	40 - 580 MB	NPU

License

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

References

• Mask R-CNN
• 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.