Segmentation Transformer: Object-Contextual Representations for Semantic Segmentation
Paper • 1909.11065 • Published
HRNet-W48-OCR: Optimized for Qualcomm Devices
HRNet-W48-OCR is a machine learning model that can segment images from the Cityscape dataset. It has lightweight and hardware-efficient operations and thus delivers significant speedup on diverse hardware platforms
This is based on the implementation of HRNet-W48-OCR 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 1.27.1 | Download |
| ONNX | w8a16 | Universal | QAIRT 2.50, ONNX Runtime 1.27.1 | Download |
For more device-specific assets and performance metrics, visit HRNet-W48-OCR 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 HRNet-W48-OCR on GitHub for usage instructions.
Model Details
Model Type: Model_use_case.semantic_segmentation
Model Stats:
- Input resolution: 2048x1024
- Model checkpoint: hrnet_ocr_cs_8162_torch11.pth
- Model size (float): 268 MB
- Model size (w8a16): 70.3 MB
- Number of output classes: 19
- Number of parameters: 70.3M
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| HRNet-W48-OCR | ONNX | float | Snapdragon® 8 Elite Gen 5 For Galaxy Mobile | 462.282 ms | 21 - 1139 MB | NPU |
| HRNet-W48-OCR | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 639.752 ms | 12 - 1063 MB | NPU |
| HRNet-W48-OCR | ONNX | float | Snapdragon® X2 Elite | 540.091 ms | 33 - 33 MB | NPU |
| HRNet-W48-OCR | ONNX | float | Snapdragon® X Elite | 998.821 ms | 132 - 132 MB | NPU |
| HRNet-W48-OCR | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 678.269 ms | 25 - 2219 MB | NPU |
| HRNet-W48-OCR | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 1456.968 ms | 26 - 2133 MB | NPU |
| HRNet-W48-OCR | ONNX | float | Qualcomm® Dragonwing™ IQ-8275 | 1142.534 ms | 24 - 52 MB | NPU |
| HRNet-W48-OCR | ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 935.261 ms | 24 - 28 MB | NPU |
| HRNet-W48-OCR | ONNX | float | Qualcomm® QCS8450 | 1456.968 ms | 26 - 2133 MB | NPU |
| HRNet-W48-OCR | ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 1135.83 ms | 24 - 51 MB | NPU |
| HRNet-W48-OCR | ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 998.821 ms | 132 - 132 MB | NPU |
| HRNet-W48-OCR | ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 639.752 ms | 12 - 1063 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Snapdragon® 8 Elite Gen 5 For Galaxy Mobile | 413.187 ms | 13 - 2014 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Snapdragon® 8 Elite For Galaxy Mobile | 532.471 ms | 6 - 1801 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Snapdragon® X2 Elite | 452.128 ms | 17 - 17 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Snapdragon® X Elite | 748.732 ms | 79 - 79 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Snapdragon® 8 Gen 3 Mobile | 631.084 ms | 0 - 2608 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Snapdragon® 8 Gen 1 Mobile | 1018.392 ms | 7 - 2663 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Qualcomm® Dragonwing™ QCS6490 | 2944.892 ms | 0 - 16 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-8275 | 740.31 ms | 5 - 21 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 832.832 ms | 0 - 86 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Qualcomm® QCS8450 | 1018.392 ms | 7 - 2663 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-9075 | 898.195 ms | 9 - 24 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-X7181 | 748.732 ms | 79 - 79 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-7790 | 1073.522 ms | 11 - 2296 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-8750 | 532.471 ms | 6 - 1801 MB | NPU |
| HRNet-W48-OCR | ONNX | w8a16 | Snapdragon® 7 Gen 4 Mobile | 1073.522 ms | 11 - 2296 MB | NPU |
License
- The license for the original implementation of HRNet-W48-OCR can be found here.
References
- Segmentation Transformer: Object-Contextual Representations 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.