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
QNN_DLC	float	Universal	QAIRT 2.43	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:

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

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
MaskRCNNProposalGenerator	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	56.939 ms	7 - 1326 MB	NPU
MaskRCNNProposalGenerator	QNN_DLC	float	Snapdragon® X2 Elite	58.441 ms	7 - 7 MB	NPU
MaskRCNNProposalGenerator	QNN_DLC	float	Snapdragon® X Elite	139.008 ms	7 - 7 MB	NPU
MaskRCNNProposalGenerator	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	105.679 ms	7 - 1393 MB	NPU
MaskRCNNProposalGenerator	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	415.448 ms	2 - 1195 MB	NPU
MaskRCNNProposalGenerator	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	148.486 ms	7 - 751 MB	NPU
MaskRCNNProposalGenerator	QNN_DLC	float	Qualcomm® SA8775P	167.193 ms	1 - 1195 MB	NPU
MaskRCNNProposalGenerator	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	210.601 ms	8 - 1353 MB	NPU
MaskRCNNProposalGenerator	QNN_DLC	float	Qualcomm® SA7255P	415.448 ms	2 - 1195 MB	NPU
MaskRCNNProposalGenerator	QNN_DLC	float	Qualcomm® SA8295P	168.736 ms	0 - 1139 MB	NPU
MaskRCNNProposalGenerator	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	71.498 ms	7 - 1304 MB	NPU
MaskRCNNROIHead	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	93.517 ms	51 - 854 MB	NPU
MaskRCNNROIHead	QNN_DLC	float	Snapdragon® X2 Elite	99.886 ms	52 - 52 MB	NPU
MaskRCNNROIHead	QNN_DLC	float	Snapdragon® X Elite	235.822 ms	52 - 52 MB	NPU
MaskRCNNROIHead	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	180.185 ms	49 - 929 MB	NPU
MaskRCNNROIHead	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	596.745 ms	49 - 849 MB	NPU
MaskRCNNROIHead	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	242.165 ms	52 - 54 MB	NPU
MaskRCNNROIHead	QNN_DLC	float	Qualcomm® SA8775P	1129.878 ms	40 - 841 MB	NPU
MaskRCNNROIHead	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	326.792 ms	39 - 940 MB	NPU
MaskRCNNROIHead	QNN_DLC	float	Qualcomm® SA7255P	596.745 ms	49 - 849 MB	NPU
MaskRCNNROIHead	QNN_DLC	float	Qualcomm® SA8295P	302.095 ms	49 - 974 MB	NPU
MaskRCNNROIHead	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	126.714 ms	34 - 824 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.