README.md

---
library_name: pytorch
license: other
tags:
- real_time
- android
pipeline_tag: image-segmentation

---

![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/bisenet/web-assets/model_demo.png)

# BiseNet: Optimized for Qualcomm Devices

BiSeNet (Bilateral Segmentation Network) is a novel architecture designed for real-time semantic segmentation. It addresses the challenge of balancing spatial resolution and receptive field by employing a Spatial Path to preserve high-resolution features and a context path to capture sufficient receptive field.

This is based on the implementation of BiseNet found [here](https://github.com/ooooverflow/BiSeNet).
This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the [Qualcomm® AI Hub Models](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/models/bisenet) 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 2.37, ONNX Runtime 1.23.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/bisenet/releases/v0.46.0/bisenet-onnx-float.zip)
| ONNX | w8a8 | Universal | QAIRT 2.37, ONNX Runtime 1.23.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/bisenet/releases/v0.46.0/bisenet-onnx-w8a8.zip)
| QNN_DLC | float | Universal | QAIRT 2.42 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/bisenet/releases/v0.46.0/bisenet-qnn_dlc-float.zip)
| QNN_DLC | w8a8 | Universal | QAIRT 2.42 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/bisenet/releases/v0.46.0/bisenet-qnn_dlc-w8a8.zip)
| TFLITE | float | Universal | QAIRT 2.42, TFLite 2.17.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/bisenet/releases/v0.46.0/bisenet-tflite-float.zip)
| TFLITE | w8a8 | Universal | QAIRT 2.42, TFLite 2.17.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/bisenet/releases/v0.46.0/bisenet-tflite-w8a8.zip)

For more device-specific assets and performance metrics, visit **[BiseNet on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/bisenet)**.


### Option 2: Export with Custom Configurations

Use the [Qualcomm® AI Hub Models](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/models/bisenet) 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 [BiseNet on GitHub](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/models/bisenet) for usage instructions.

## Model Details

**Model Type:** Model_use_case.semantic_segmentation

**Model Stats:**
- Model checkpoint: best_dice_loss_miou_0.655.pth
- Inference latency: RealTime
- Input resolution: 720x960
- Number of parameters: 12.0M
- Model size (float): 45.7 MB

## Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit
|---|---|---|---|---|---|---
| BiseNet | ONNX | float | Snapdragon® X Elite | 31.468 ms | 66 - 66 MB | NPU
| BiseNet | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 26.195 ms | 73 - 270 MB | NPU
| BiseNet | ONNX | float | Qualcomm® QCS8550 (Proxy) | 32.87 ms | 63 - 86 MB | NPU
| BiseNet | ONNX | float | Qualcomm® QCS9075 | 51.221 ms | 8 - 11 MB | NPU
| BiseNet | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 19.854 ms | 71 - 211 MB | NPU
| BiseNet | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 15.173 ms | 56 - 204 MB | NPU
| BiseNet | ONNX | w8a8 | Snapdragon® X Elite | 8.703 ms | 19 - 19 MB | NPU
| BiseNet | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 5.962 ms | 18 - 210 MB | NPU
| BiseNet | ONNX | w8a8 | Qualcomm® QCS6490 | 236.082 ms | 223 - 236 MB | CPU
| BiseNet | ONNX | w8a8 | Qualcomm® QCS8550 (Proxy) | 8.607 ms | 16 - 45 MB | NPU
| BiseNet | ONNX | w8a8 | Qualcomm® QCS9075 | 10.345 ms | 18 - 21 MB | NPU
| BiseNet | ONNX | w8a8 | Qualcomm® QCM6690 | 232.957 ms | 132 - 139 MB | CPU
| BiseNet | ONNX | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 4.798 ms | 17 - 164 MB | NPU
| BiseNet | ONNX | w8a8 | Snapdragon® 7 Gen 4 Mobile | 204.645 ms | 212 - 219 MB | CPU
| BiseNet | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 3.743 ms | 0 - 151 MB | NPU
| BiseNet | QNN_DLC | float | Snapdragon® X Elite | 28.927 ms | 8 - 8 MB | NPU
| BiseNet | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 20.22 ms | 8 - 285 MB | NPU
| BiseNet | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 107.749 ms | 2 - 194 MB | NPU
| BiseNet | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 28.517 ms | 8 - 10 MB | NPU
| BiseNet | QNN_DLC | float | Qualcomm® SA8775P | 38.769 ms | 1 - 188 MB | NPU
| BiseNet | QNN_DLC | float | Qualcomm® QCS9075 | 55.43 ms | 8 - 49 MB | NPU
| BiseNet | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 59.857 ms | 8 - 277 MB | NPU
| BiseNet | QNN_DLC | float | Qualcomm® SA7255P | 107.749 ms | 2 - 194 MB | NPU
| BiseNet | QNN_DLC | float | Qualcomm® SA8295P | 44.137 ms | 0 - 213 MB | NPU
| BiseNet | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 15.012 ms | 8 - 262 MB | NPU
| BiseNet | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 11.918 ms | 6 - 284 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Snapdragon® X Elite | 10.122 ms | 2 - 2 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 6.747 ms | 2 - 233 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Qualcomm® QCS6490 | 40.586 ms | 2 - 14 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Qualcomm® QCS8275 (Proxy) | 20.155 ms | 2 - 182 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Qualcomm® QCS8550 (Proxy) | 9.474 ms | 2 - 4 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Qualcomm® SA8775P | 10.25 ms | 2 - 183 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Qualcomm® QCS9075 | 13.068 ms | 2 - 14 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Qualcomm® QCM6690 | 90.291 ms | 2 - 206 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Qualcomm® QCS8450 (Proxy) | 16.163 ms | 2 - 231 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Qualcomm® SA7255P | 20.155 ms | 2 - 182 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Qualcomm® SA8295P | 12.588 ms | 2 - 185 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 5.175 ms | 2 - 192 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 13.403 ms | 2 - 198 MB | NPU
| BiseNet | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 4.278 ms | 2 - 193 MB | NPU
| BiseNet | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 20.372 ms | 31 - 310 MB | NPU
| BiseNet | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 105.404 ms | 32 - 247 MB | NPU
| BiseNet | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 27.811 ms | 32 - 34 MB | NPU
| BiseNet | TFLITE | float | Qualcomm® SA8775P | 37.795 ms | 32 - 246 MB | NPU
| BiseNet | TFLITE | float | Qualcomm® QCS9075 | 54.488 ms | 0 - 66 MB | NPU
| BiseNet | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 59.939 ms | 32 - 307 MB | NPU
| BiseNet | TFLITE | float | Qualcomm® SA7255P | 105.404 ms | 32 - 247 MB | NPU
| BiseNet | TFLITE | float | Qualcomm® SA8295P | 44.229 ms | 23 - 237 MB | NPU
| BiseNet | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 15.177 ms | 30 - 288 MB | NPU
| BiseNet | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 11.908 ms | 30 - 309 MB | NPU
| BiseNet | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 8.658 ms | 7 - 240 MB | NPU
| BiseNet | TFLITE | w8a8 | Qualcomm® QCS6490 | 47.105 ms | 6 - 30 MB | NPU
| BiseNet | TFLITE | w8a8 | Qualcomm® QCS8275 (Proxy) | 20.753 ms | 0 - 182 MB | NPU
| BiseNet | TFLITE | w8a8 | Qualcomm® QCS8550 (Proxy) | 12.154 ms | 0 - 110 MB | NPU
| BiseNet | TFLITE | w8a8 | Qualcomm® SA8775P | 12.707 ms | 0 - 183 MB | NPU
| BiseNet | TFLITE | w8a8 | Qualcomm® QCS9075 | 13.12 ms | 8 - 32 MB | NPU
| BiseNet | TFLITE | w8a8 | Qualcomm® QCM6690 | 101.835 ms | 6 - 209 MB | NPU
| BiseNet | TFLITE | w8a8 | Qualcomm® QCS8450 (Proxy) | 16.289 ms | 8 - 238 MB | NPU
| BiseNet | TFLITE | w8a8 | Qualcomm® SA7255P | 20.753 ms | 0 - 182 MB | NPU
| BiseNet | TFLITE | w8a8 | Qualcomm® SA8295P | 15.152 ms | 8 - 194 MB | NPU
| BiseNet | TFLITE | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 6.588 ms | 6 - 200 MB | NPU
| BiseNet | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 15.957 ms | 0 - 199 MB | NPU
| BiseNet | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 5.515 ms | 6 - 199 MB | NPU

## License
* The license for the original implementation of BiseNet can be found
  [here](https://github.com/ooooverflow/BiSeNet/pull/45/files).

## References
* [BiSeNet Bilateral Segmentation Network for Real-time Semantic Segmentation](https://arxiv.org/abs/1808.00897)
* [Source Model Implementation](https://github.com/ooooverflow/BiSeNet)

## 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).