| --- |
| library_name: pytorch |
| license: other |
| tags: |
| - real_time |
| - android |
| pipeline_tag: image-segmentation |
|
|
| --- |
| |
|  |
|
|
| # PidNet: Optimized for Qualcomm Devices |
|
|
| PIDNet (Proportional-Integral-Derivative Network) is a real-time semantic segmentation model based on PID controllers |
|
|
| This is based on the implementation of PidNet found [here](https://github.com/XuJiacong/PIDNet). |
| This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/main/qai_hub_models/models/pidnet) library to export with custom configurations. More details on model performance across various devices, can be found [here](#performance-summary). |
|
|
| Qualcomm AI Hub Models uses [Qualcomm AI Hub Workbench](https://workbench.aihub.qualcomm.com) to compile, profile, and evaluate this model. [Sign up](https://myaccount.qualcomm.com/signup) 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](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/pidnet/releases/v0.49.1/pidnet-onnx-float.zip) |
| | ONNX | w8a8 | Universal | QAIRT 2.42, ONNX Runtime 1.24.1 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/pidnet/releases/v0.49.1/pidnet-onnx-w8a8.zip) |
| | QNN_DLC | float | Universal | QAIRT 2.43 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/pidnet/releases/v0.49.1/pidnet-qnn_dlc-float.zip) |
| | QNN_DLC | w8a8 | Universal | QAIRT 2.43 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/pidnet/releases/v0.49.1/pidnet-qnn_dlc-w8a8.zip) |
| | TFLITE | float | Universal | QAIRT 2.43, TFLite 2.17.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/pidnet/releases/v0.49.1/pidnet-tflite-float.zip) |
| | TFLITE | w8a8 | Universal | QAIRT 2.43, TFLite 2.17.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/pidnet/releases/v0.49.1/pidnet-tflite-w8a8.zip) |
|
|
| For more device-specific assets and performance metrics, visit **[PidNet on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/pidnet)**. |
|
|
|
|
| ### Option 2: Export with Custom Configurations |
|
|
| Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/main/qai_hub_models/models/pidnet) 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 [PidNet on GitHub](https://github.com/qualcomm/ai-hub-models/blob/main/qai_hub_models/models/pidnet) for usage instructions. |
|
|
| ## Model Details |
|
|
| **Model Type:** Model_use_case.semantic_segmentation |
| |
| **Model Stats:** |
| - Model checkpoint: PIDNet_S_Cityscapes_val.pt |
| - Inference latency: RealTime |
| - Input resolution: 1024x2048 |
| - Number of output classes: 19 |
| - Number of parameters: 8.06M |
| - Model size (float): 29.1 MB |
| - Model size (w8a8): 8.02 MB |
|
|
| ## Performance Summary |
| | Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
| |---|---|---|---|---|---|--- |
| | PidNet | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 12.188 ms | 30 - 298 MB | NPU |
| | PidNet | ONNX | float | Snapdragon® X2 Elite | 13.225 ms | 22 - 22 MB | NPU |
| | PidNet | ONNX | float | Snapdragon® X Elite | 32.46 ms | 24 - 24 MB | NPU |
| | PidNet | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 23.02 ms | 1 - 318 MB | NPU |
| | PidNet | ONNX | float | Qualcomm® QCS8550 (Proxy) | 33.333 ms | 24 - 27 MB | NPU |
| | PidNet | ONNX | float | Qualcomm® QCS9075 | 47.508 ms | 24 - 51 MB | NPU |
| | PidNet | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 16.566 ms | 5 - 225 MB | NPU |
| | PidNet | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 44.113 ms | 71 - 291 MB | NPU |
| | PidNet | ONNX | w8a8 | Snapdragon® X2 Elite | 45.785 ms | 134 - 134 MB | NPU |
| | PidNet | ONNX | w8a8 | Snapdragon® X Elite | 89.05 ms | 133 - 133 MB | NPU |
| | PidNet | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 45.845 ms | 110 - 371 MB | NPU |
| | PidNet | ONNX | w8a8 | Qualcomm® QCS6490 | 391.518 ms | 197 - 215 MB | CPU |
| | PidNet | ONNX | w8a8 | Qualcomm® QCS8550 (Proxy) | 59.651 ms | 102 - 106 MB | NPU |
| | PidNet | ONNX | w8a8 | Qualcomm® QCS9075 | 66.387 ms | 105 - 108 MB | NPU |
| | PidNet | ONNX | w8a8 | Qualcomm® QCM6690 | 346.885 ms | 198 - 207 MB | CPU |
| | PidNet | ONNX | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 43.647 ms | 99 - 305 MB | NPU |
| | PidNet | ONNX | w8a8 | Snapdragon® 7 Gen 4 Mobile | 326.982 ms | 223 - 233 MB | CPU |
| | PidNet | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 11.958 ms | 12 - 284 MB | NPU |
| | PidNet | QNN_DLC | float | Snapdragon® X2 Elite | 13.5 ms | 24 - 24 MB | NPU |
| | PidNet | QNN_DLC | float | Snapdragon® X Elite | 38.855 ms | 24 - 24 MB | NPU |
| | PidNet | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 25.603 ms | 24 - 344 MB | NPU |
| | PidNet | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 118.292 ms | 24 - 246 MB | NPU |
| | PidNet | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 37.634 ms | 24 - 26 MB | NPU |
| | PidNet | QNN_DLC | float | Qualcomm® SA8775P | 46.74 ms | 24 - 247 MB | NPU |
| | PidNet | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 76.339 ms | 5 - 335 MB | NPU |
| | PidNet | QNN_DLC | float | Qualcomm® SA7255P | 118.292 ms | 24 - 246 MB | NPU |
| | PidNet | QNN_DLC | float | Qualcomm® SA8295P | 51.692 ms | 24 - 261 MB | NPU |
| | PidNet | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 18.325 ms | 10 - 255 MB | NPU |
| | PidNet | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 47.609 ms | 6 - 272 MB | NPU |
| | PidNet | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 48.32 ms | 6 - 6 MB | NPU |
| | PidNet | QNN_DLC | w8a8 | Snapdragon® X Elite | 60.394 ms | 6 - 6 MB | NPU |
| | PidNet | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 43.285 ms | 6 - 273 MB | NPU |
| | PidNet | QNN_DLC | w8a8 | Qualcomm® QCS8275 (Proxy) | 111.157 ms | 6 - 218 MB | NPU |
| | PidNet | QNN_DLC | w8a8 | Qualcomm® QCS8550 (Proxy) | 57.6 ms | 6 - 161 MB | NPU |
| | PidNet | QNN_DLC | w8a8 | Qualcomm® SA8775P | 58.416 ms | 6 - 218 MB | NPU |
| | PidNet | QNN_DLC | w8a8 | Qualcomm® QCS9075 | 61.338 ms | 6 - 14 MB | NPU |
| | PidNet | QNN_DLC | w8a8 | Qualcomm® QCS8450 (Proxy) | 64.153 ms | 6 - 271 MB | NPU |
| | PidNet | QNN_DLC | w8a8 | Qualcomm® SA7255P | 111.157 ms | 6 - 218 MB | NPU |
| | PidNet | QNN_DLC | w8a8 | Qualcomm® SA8295P | 66.448 ms | 6 - 220 MB | NPU |
| | PidNet | QNN_DLC | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 41.4 ms | 6 - 240 MB | NPU |
| | PidNet | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 12.644 ms | 2 - 272 MB | NPU |
| | PidNet | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 24.886 ms | 0 - 334 MB | NPU |
| | PidNet | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 117.098 ms | 3 - 231 MB | NPU |
| | PidNet | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 36.18 ms | 2 - 5 MB | NPU |
| | PidNet | TFLITE | float | Qualcomm® SA8775P | 45.314 ms | 2 - 231 MB | NPU |
| | PidNet | TFLITE | float | Qualcomm® QCS9075 | 60.142 ms | 0 - 45 MB | NPU |
| | PidNet | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 75.223 ms | 3 - 350 MB | NPU |
| | PidNet | TFLITE | float | Qualcomm® SA7255P | 117.098 ms | 3 - 231 MB | NPU |
| | PidNet | TFLITE | float | Qualcomm® SA8295P | 50.603 ms | 3 - 249 MB | NPU |
| | PidNet | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 18.63 ms | 1 - 254 MB | NPU |
| | PidNet | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 40.789 ms | 1 - 265 MB | NPU |
| | PidNet | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 37.924 ms | 1 - 268 MB | NPU |
| | PidNet | TFLITE | w8a8 | Qualcomm® QCS6490 | 203.117 ms | 3 - 73 MB | NPU |
| | PidNet | TFLITE | w8a8 | Qualcomm® QCS8275 (Proxy) | 98.248 ms | 1 - 211 MB | NPU |
| | PidNet | TFLITE | w8a8 | Qualcomm® QCS8550 (Proxy) | 50.296 ms | 1 - 3 MB | NPU |
| | PidNet | TFLITE | w8a8 | Qualcomm® SA8775P | 50.98 ms | 0 - 212 MB | NPU |
| | PidNet | TFLITE | w8a8 | Qualcomm® QCS9075 | 52.952 ms | 1 - 17 MB | NPU |
| | PidNet | TFLITE | w8a8 | Qualcomm® QCM6690 | 242.39 ms | 3 - 197 MB | NPU |
| | PidNet | TFLITE | w8a8 | Qualcomm® QCS8450 (Proxy) | 58.405 ms | 1 - 266 MB | NPU |
| | PidNet | TFLITE | w8a8 | Qualcomm® SA7255P | 98.248 ms | 1 - 211 MB | NPU |
| | PidNet | TFLITE | w8a8 | Qualcomm® SA8295P | 57.929 ms | 1 - 214 MB | NPU |
| | PidNet | TFLITE | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 38.294 ms | 0 - 236 MB | NPU |
| | PidNet | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 67.211 ms | 3 - 223 MB | NPU |
| |
| ## License |
| * The license for the original implementation of PidNet can be found |
| [here](https://github.com/XuJiacong/PIDNet/blob/main/LICENSE). |
| |
| ## References |
| * [PIDNet A Real-time Semantic Segmentation Network Inspired from PID Controller Segmentation of Road Scenes](https://arxiv.org/abs/2206.02066) |
| * [Source Model Implementation](https://github.com/XuJiacong/PIDNet) |
| |
| ## Community |
| * Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI. |
| * For questions or feedback please [reach out to us](mailto:ai-hub-support@qti.qualcomm.com). |
| |
