Rethinking Atrous Convolution for Semantic Image Segmentation
Paper
• 1706.05587 • Published
DeepLabV3-Plus-MobileNet: Optimized for Qualcomm Devices
DeepLabV3 is designed for semantic segmentation at multiple scales, trained on the various datasets. It uses MobileNet as a backbone.
This is based on the implementation of DeepLabV3-Plus-MobileNet 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 2.42, ONNX Runtime 1.24.1 | Download |
| ONNX | w8a16 | Universal | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| ONNX | w8a8 | Universal | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| QNN_DLC | float | Universal | QAIRT 2.43 | Download |
| QNN_DLC | w8a16 | Universal | QAIRT 2.43 | Download |
| QNN_DLC | w8a8 | Universal | QAIRT 2.43 | Download |
| TFLITE | float | Universal | QAIRT 2.43, TFLite 2.17.0 | Download |
| TFLITE | w8a8 | Universal | QAIRT 2.43, TFLite 2.17.0 | Download |
For more device-specific assets and performance metrics, visit DeepLabV3-Plus-MobileNet 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 DeepLabV3-Plus-MobileNet on GitHub for usage instructions.
Model Details
Model Type: Model_use_case.semantic_segmentation
Model Stats:
- Model checkpoint: VOC2012
- Input resolution: 513x513
- Number of output classes: 21
- Number of parameters: 5.80M
- Model size (float): 22.2 MB
- Model size (w8a16): 6.67 MB
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| DeepLabV3-Plus-MobileNet | ONNX | float | Snapdragon® X2 Elite | 5.383 ms | 10 - 10 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | float | Snapdragon® X Elite | 11.032 ms | 10 - 10 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 7.204 ms | 4 - 215 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | float | Qualcomm® QCS8550 (Proxy) | 10.345 ms | 3 - 5 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | float | Qualcomm® QCS9075 | 18.027 ms | 3 - 6 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 5.894 ms | 1 - 170 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 4.303 ms | 1 - 175 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a16 | Snapdragon® X2 Elite | 3.716 ms | 7 - 7 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a16 | Snapdragon® X Elite | 7.688 ms | 5 - 5 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a16 | Snapdragon® 8 Gen 3 Mobile | 5.097 ms | 2 - 219 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a16 | Qualcomm® QCS6490 | 1221.485 ms | 89 - 92 MB | CPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a16 | Qualcomm® QCS8550 (Proxy) | 7.073 ms | 2 - 5 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a16 | Qualcomm® QCS9075 | 8.616 ms | 2 - 4 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a16 | Qualcomm® QCM6690 | 589.325 ms | 107 - 115 MB | CPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a16 | Snapdragon® 8 Elite For Galaxy Mobile | 3.937 ms | 0 - 184 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a16 | Snapdragon® 7 Gen 4 Mobile | 585.899 ms | 92 - 101 MB | CPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 2.962 ms | 0 - 210 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a8 | Snapdragon® X2 Elite | 1.672 ms | 7 - 7 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a8 | Snapdragon® X Elite | 3.927 ms | 5 - 5 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 2.374 ms | 0 - 203 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a8 | Qualcomm® QCS8550 (Proxy) | 3.537 ms | 0 - 3 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a8 | Qualcomm® QCS9075 | 4.461 ms | 1 - 4 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 1.904 ms | 0 - 176 MB | NPU |
| DeepLabV3-Plus-MobileNet | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 1.462 ms | 0 - 179 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | float | Snapdragon® X2 Elite | 5.613 ms | 3 - 3 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | float | Snapdragon® X Elite | 12.228 ms | 3 - 3 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 8.047 ms | 3 - 213 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 58.591 ms | 1 - 167 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 11.508 ms | 3 - 5 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | float | Qualcomm® SA8775P | 17.521 ms | 0 - 171 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | float | Qualcomm® QCS9075 | 20.252 ms | 5 - 10 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 18.981 ms | 1 - 211 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | float | Qualcomm® SA7255P | 58.591 ms | 1 - 167 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | float | Qualcomm® SA8295P | 19.058 ms | 0 - 174 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 6.25 ms | 3 - 199 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 4.295 ms | 3 - 195 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Snapdragon® X2 Elite | 4.163 ms | 2 - 2 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Snapdragon® X Elite | 9.147 ms | 2 - 2 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Snapdragon® 8 Gen 3 Mobile | 6.112 ms | 0 - 218 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Qualcomm® QCS6490 | 33.548 ms | 1 - 4 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Qualcomm® QCS8275 (Proxy) | 21.941 ms | 2 - 182 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Qualcomm® QCS8550 (Proxy) | 8.473 ms | 2 - 4 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Qualcomm® SA8775P | 40.099 ms | 2 - 181 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Qualcomm® QCS9075 | 9.797 ms | 1 - 4 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Qualcomm® QCM6690 | 100.543 ms | 2 - 233 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Qualcomm® QCS8450 (Proxy) | 12.062 ms | 2 - 222 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Qualcomm® SA7255P | 21.941 ms | 2 - 182 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Qualcomm® SA8295P | 13.74 ms | 2 - 197 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Snapdragon® 8 Elite For Galaxy Mobile | 4.328 ms | 2 - 181 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Snapdragon® 7 Gen 4 Mobile | 12.099 ms | 2 - 180 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 3.073 ms | 2 - 197 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 2.055 ms | 1 - 1 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Snapdragon® X Elite | 4.533 ms | 1 - 1 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 2.9 ms | 0 - 196 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Qualcomm® QCS6490 | 15.752 ms | 3 - 5 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Qualcomm® QCS8275 (Proxy) | 10.789 ms | 1 - 170 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Qualcomm® QCS8550 (Proxy) | 4.123 ms | 1 - 2 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Qualcomm® SA8775P | 4.62 ms | 1 - 172 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Qualcomm® QCS9075 | 5.015 ms | 1 - 3 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Qualcomm® QCM6690 | 47.555 ms | 1 - 202 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Qualcomm® QCS8450 (Proxy) | 6.421 ms | 1 - 198 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Qualcomm® SA7255P | 10.789 ms | 1 - 170 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Qualcomm® SA8295P | 6.412 ms | 1 - 167 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 2.14 ms | 1 - 171 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 6.2 ms | 1 - 175 MB | NPU |
| DeepLabV3-Plus-MobileNet | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 1.507 ms | 1 - 175 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 8.052 ms | 0 - 221 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 58.6 ms | 0 - 170 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 11.505 ms | 0 - 3 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | float | Qualcomm® SA8775P | 17.571 ms | 0 - 174 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | float | Qualcomm® QCS9075 | 19.652 ms | 0 - 18 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 18.977 ms | 0 - 217 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | float | Qualcomm® SA7255P | 58.6 ms | 0 - 170 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | float | Qualcomm® SA8295P | 19.131 ms | 0 - 174 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 6.244 ms | 0 - 199 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 4.294 ms | 0 - 193 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 2.955 ms | 0 - 207 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | w8a8 | Qualcomm® QCS6490 | 15.843 ms | 0 - 11 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | w8a8 | Qualcomm® QCS8275 (Proxy) | 10.875 ms | 0 - 173 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | w8a8 | Qualcomm® QCS8550 (Proxy) | 4.31 ms | 0 - 10 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | w8a8 | Qualcomm® SA8775P | 4.658 ms | 0 - 175 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | w8a8 | Qualcomm® QCS9075 | 5.166 ms | 0 - 10 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | w8a8 | Qualcomm® QCM6690 | 52.604 ms | 0 - 193 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | w8a8 | Qualcomm® QCS8450 (Proxy) | 5.75 ms | 0 - 205 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | w8a8 | Qualcomm® SA7255P | 10.875 ms | 0 - 173 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | w8a8 | Qualcomm® SA8295P | 6.241 ms | 0 - 172 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 2.093 ms | 0 - 177 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 6.365 ms | 0 - 178 MB | NPU |
| DeepLabV3-Plus-MobileNet | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 1.523 ms | 0 - 182 MB | NPU |
License
- The license for the original implementation of DeepLabV3-Plus-MobileNet can be found here.
References
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.