README.md

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

---

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

# Unet-Segmentation: Optimized for Qualcomm Devices

UNet is a machine learning model that produces a segmentation mask for an image. The most basic use case will label each pixel in the image as being in the foreground or the background. More advanced usage will assign a class label to each pixel. This version of the model was trained on the data from Kaggle's Carvana Image Masking Challenge (see https://www.kaggle.com/c/carvana-image-masking-challenge) and is used for vehicle segmentation.

This is based on the implementation of Unet-Segmentation found [here](https://github.com/milesial/Pytorch-UNet).
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/unet_segmentation) 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/unet_segmentation/releases/v0.48.0/unet_segmentation-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/unet_segmentation/releases/v0.48.0/unet_segmentation-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/unet_segmentation/releases/v0.48.0/unet_segmentation-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/unet_segmentation/releases/v0.48.0/unet_segmentation-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/unet_segmentation/releases/v0.48.0/unet_segmentation-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/unet_segmentation/releases/v0.48.0/unet_segmentation-tflite-w8a8.zip)

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


### 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/unet_segmentation) 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 [Unet-Segmentation on GitHub](https://github.com/qualcomm/ai-hub-models/blob/main/qai_hub_models/models/unet_segmentation) for usage instructions.

## Model Details

**Model Type:** Model_use_case.semantic_segmentation

**Model Stats:**
- Model checkpoint: unet_carvana_scale1.0_epoch2
- Input resolution: 224x224
- Number of output classes: 2 (foreground / background)
- Number of parameters: 31.0M
- Model size (float): 118 MB
- Model size (w8a8): 29.8 MB

## Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit
|---|---|---|---|---|---|---
| Unet-Segmentation | ONNX | float | Snapdragon® X2 Elite | 75.129 ms | 53 - 53 MB | NPU
| Unet-Segmentation | ONNX | float | Snapdragon® X Elite | 139.574 ms | 53 - 53 MB | NPU
| Unet-Segmentation | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 109.506 ms | 25 - 562 MB | NPU
| Unet-Segmentation | ONNX | float | Qualcomm® QCS8550 (Proxy) | 144.021 ms | 0 - 58 MB | NPU
| Unet-Segmentation | ONNX | float | Qualcomm® QCS9075 | 254.662 ms | 9 - 21 MB | NPU
| Unet-Segmentation | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 89.057 ms | 14 - 330 MB | NPU
| Unet-Segmentation | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 66.983 ms | 4 - 327 MB | NPU
| Unet-Segmentation | ONNX | w8a8 | Snapdragon® X2 Elite | 20.038 ms | 29 - 29 MB | NPU
| Unet-Segmentation | ONNX | w8a8 | Snapdragon® X Elite | 39.087 ms | 29 - 29 MB | NPU
| Unet-Segmentation | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 30.409 ms | 6 - 338 MB | NPU
| Unet-Segmentation | ONNX | w8a8 | Qualcomm® QCS6490 | 4677.804 ms | 943 - 1000 MB | CPU
| Unet-Segmentation | ONNX | w8a8 | Qualcomm® QCS8550 (Proxy) | 39.501 ms | 0 - 12 MB | NPU
| Unet-Segmentation | ONNX | w8a8 | Qualcomm® QCS9075 | 35.587 ms | 4 - 7 MB | NPU
| Unet-Segmentation | ONNX | w8a8 | Qualcomm® QCM6690 | 4143.656 ms | 835 - 842 MB | CPU
| Unet-Segmentation | ONNX | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 24.647 ms | 3 - 189 MB | NPU
| Unet-Segmentation | ONNX | w8a8 | Snapdragon® 7 Gen 4 Mobile | 3886.063 ms | 833 - 840 MB | CPU
| Unet-Segmentation | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 16.357 ms | 6 - 189 MB | NPU
| Unet-Segmentation | QNN_DLC | float | Snapdragon® X2 Elite | 72.06 ms | 9 - 9 MB | NPU
| Unet-Segmentation | QNN_DLC | float | Snapdragon® X Elite | 132.493 ms | 9 - 9 MB | NPU
| Unet-Segmentation | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 101.922 ms | 9 - 542 MB | NPU
| Unet-Segmentation | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 953.451 ms | 0 - 323 MB | NPU
| Unet-Segmentation | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 137.283 ms | 10 - 12 MB | NPU
| Unet-Segmentation | QNN_DLC | float | Qualcomm® SA8775P | 240.471 ms | 0 - 323 MB | NPU
| Unet-Segmentation | QNN_DLC | float | Qualcomm® QCS9075 | 248.401 ms | 9 - 27 MB | NPU
| Unet-Segmentation | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 277.158 ms | 9 - 548 MB | NPU
| Unet-Segmentation | QNN_DLC | float | Qualcomm® SA7255P | 953.451 ms | 0 - 323 MB | NPU
| Unet-Segmentation | QNN_DLC | float | Qualcomm® SA8295P | 274.522 ms | 0 - 322 MB | NPU
| Unet-Segmentation | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 82.983 ms | 0 - 332 MB | NPU
| Unet-Segmentation | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 63.074 ms | 9 - 350 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 18.867 ms | 2 - 2 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Snapdragon® X Elite | 35.686 ms | 2 - 2 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 26.305 ms | 2 - 321 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Qualcomm® QCS6490 | 267.84 ms | 2 - 8 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Qualcomm® QCS8275 (Proxy) | 121.511 ms | 1 - 180 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Qualcomm® QCS8550 (Proxy) | 34.694 ms | 2 - 4 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Qualcomm® SA8775P | 32.223 ms | 1 - 180 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Qualcomm® QCS9075 | 34.299 ms | 2 - 8 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Qualcomm® QCM6690 | 1243.995 ms | 2 - 521 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Qualcomm® QCS8450 (Proxy) | 60.657 ms | 3 - 321 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Qualcomm® SA7255P | 121.511 ms | 1 - 180 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Qualcomm® SA8295P | 63.73 ms | 0 - 180 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 21.615 ms | 2 - 190 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 78.746 ms | 2 - 268 MB | NPU
| Unet-Segmentation | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 15.738 ms | 2 - 198 MB | NPU
| Unet-Segmentation | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 103.433 ms | 6 - 543 MB | NPU
| Unet-Segmentation | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 953.42 ms | 0 - 325 MB | NPU
| Unet-Segmentation | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 136.873 ms | 6 - 308 MB | NPU
| Unet-Segmentation | TFLITE | float | Qualcomm® SA8775P | 1126.415 ms | 5 - 330 MB | NPU
| Unet-Segmentation | TFLITE | float | Qualcomm® QCS9075 | 248.066 ms | 0 - 80 MB | NPU
| Unet-Segmentation | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 278.634 ms | 7 - 551 MB | NPU
| Unet-Segmentation | TFLITE | float | Qualcomm® SA7255P | 953.42 ms | 0 - 325 MB | NPU
| Unet-Segmentation | TFLITE | float | Qualcomm® SA8295P | 274.503 ms | 0 - 322 MB | NPU
| Unet-Segmentation | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 82.529 ms | 0 - 331 MB | NPU
| Unet-Segmentation | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 62.584 ms | 6 - 353 MB | NPU
| Unet-Segmentation | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 26.332 ms | 14 - 333 MB | NPU
| Unet-Segmentation | TFLITE | w8a8 | Qualcomm® QCS6490 | 267.765 ms | 0 - 40 MB | NPU
| Unet-Segmentation | TFLITE | w8a8 | Qualcomm® QCS8275 (Proxy) | 121.634 ms | 2 - 181 MB | NPU
| Unet-Segmentation | TFLITE | w8a8 | Qualcomm® QCS8550 (Proxy) | 32.145 ms | 2 - 623 MB | NPU
| Unet-Segmentation | TFLITE | w8a8 | Qualcomm® SA8775P | 32.24 ms | 2 - 181 MB | NPU
| Unet-Segmentation | TFLITE | w8a8 | Qualcomm® QCS9075 | 34.234 ms | 0 - 37 MB | NPU
| Unet-Segmentation | TFLITE | w8a8 | Qualcomm® QCM6690 | 1238.061 ms | 0 - 519 MB | NPU
| Unet-Segmentation | TFLITE | w8a8 | Qualcomm® QCS8450 (Proxy) | 60.326 ms | 2 - 320 MB | NPU
| Unet-Segmentation | TFLITE | w8a8 | Qualcomm® SA7255P | 121.634 ms | 2 - 181 MB | NPU
| Unet-Segmentation | TFLITE | w8a8 | Qualcomm® SA8295P | 63.769 ms | 2 - 180 MB | NPU
| Unet-Segmentation | TFLITE | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 21.825 ms | 1 - 187 MB | NPU
| Unet-Segmentation | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 78.773 ms | 1 - 269 MB | NPU
| Unet-Segmentation | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 15.695 ms | 7 - 202 MB | NPU

## License
* The license for the original implementation of Unet-Segmentation can be found
  [here](https://github.com/milesial/Pytorch-UNet/blob/master/LICENSE).

## References
* [U-Net: Convolutional Networks for Biomedical Image Segmentation](https://arxiv.org/abs/1505.04597)
* [Source Model Implementation](https://github.com/milesial/Pytorch-UNet)

## 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).