Hugging Face's logo

qualcomm
/
Unet-Segmentation

Image Segmentation
PyTorch
LiteRT
backbone
bu_auto
real_time
android
Model card Files
xet
Community

Unet-Segmentation: Optimized for Mobile Deployment

Real-time segmentation optimized for mobile and edge

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 model is an implementation of Unet-Segmentation found here.

This repository provides scripts to run Unet-Segmentation on Qualcomm® devices. More details on model performance across various devices, can be found here.

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
Model Precision Device Chipset Target Runtime Inference Time (ms) Peak Memory Range (MB) Primary Compute Unit Target Model
Unet-Segmentation float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 943.905 ms 0 - 244 MB NPU Unet-Segmentation.tflite
Unet-Segmentation float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 943.946 ms 2 - 244 MB NPU Unet-Segmentation.dlc
Unet-Segmentation float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 284.741 ms 0 - 462 MB NPU Unet-Segmentation.tflite
Unet-Segmentation float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 272.515 ms 10 - 467 MB NPU Unet-Segmentation.dlc
Unet-Segmentation float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 143.04 ms 6 - 443 MB NPU Unet-Segmentation.tflite
Unet-Segmentation float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 140.251 ms 9 - 12 MB NPU Unet-Segmentation.dlc
Unet-Segmentation float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) ONNX 168.797 ms 0 - 57 MB NPU Unet-Segmentation.onnx.zip
Unet-Segmentation float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 239.592 ms 6 - 250 MB NPU Unet-Segmentation.tflite
Unet-Segmentation float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 239.576 ms 1 - 243 MB NPU Unet-Segmentation.dlc
Unet-Segmentation float SA7255P ADP Qualcomm® SA7255P TFLITE 943.905 ms 0 - 244 MB NPU Unet-Segmentation.tflite
Unet-Segmentation float SA7255P ADP Qualcomm® SA7255P QNN_DLC 943.946 ms 2 - 244 MB NPU Unet-Segmentation.dlc
Unet-Segmentation float SA8255 (Proxy) Qualcomm® SA8255P (Proxy) TFLITE 139.432 ms 6 - 217 MB NPU Unet-Segmentation.tflite
Unet-Segmentation float SA8255 (Proxy) Qualcomm® SA8255P (Proxy) QNN_DLC 136.015 ms 10 - 13 MB NPU Unet-Segmentation.dlc
Unet-Segmentation float SA8295P ADP Qualcomm® SA8295P TFLITE 273.711 ms 0 - 242 MB NPU Unet-Segmentation.tflite
Unet-Segmentation float SA8295P ADP Qualcomm® SA8295P QNN_DLC 273.712 ms 0 - 244 MB NPU Unet-Segmentation.dlc
Unet-Segmentation float SA8650 (Proxy) Qualcomm® SA8650P (Proxy) TFLITE 139.598 ms 6 - 443 MB NPU Unet-Segmentation.tflite
Unet-Segmentation float SA8650 (Proxy) Qualcomm® SA8650P (Proxy) QNN_DLC 135.14 ms 9 - 12 MB NPU Unet-Segmentation.dlc
Unet-Segmentation float SA8775P ADP Qualcomm® SA8775P TFLITE 239.592 ms 6 - 250 MB NPU Unet-Segmentation.tflite
Unet-Segmentation float SA8775P ADP Qualcomm® SA8775P QNN_DLC 239.576 ms 1 - 243 MB NPU Unet-Segmentation.dlc
Unet-Segmentation float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 103.87 ms 0 - 453 MB NPU Unet-Segmentation.tflite
Unet-Segmentation float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 103.025 ms 10 - 463 MB NPU Unet-Segmentation.dlc
Unet-Segmentation float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 116.338 ms 24 - 454 MB NPU Unet-Segmentation.onnx.zip
Unet-Segmentation float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile TFLITE 83.302 ms 6 - 253 MB NPU Unet-Segmentation.tflite
Unet-Segmentation float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile QNN_DLC 83.885 ms 9 - 257 MB NPU Unet-Segmentation.dlc
Unet-Segmentation float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile ONNX 93.426 ms 15 - 224 MB NPU Unet-Segmentation.onnx.zip
Unet-Segmentation float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile TFLITE 68.439 ms 6 - 268 MB NPU Unet-Segmentation.tflite
Unet-Segmentation float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile QNN_DLC 62.536 ms 9 - 272 MB NPU Unet-Segmentation.dlc
Unet-Segmentation float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile ONNX 70.035 ms 24 - 243 MB NPU Unet-Segmentation.onnx.zip
Unet-Segmentation float Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 133.145 ms 9 - 9 MB NPU Unet-Segmentation.dlc
Unet-Segmentation float Snapdragon X Elite CRD Snapdragon® X Elite ONNX 144.465 ms 54 - 54 MB NPU Unet-Segmentation.onnx.zip
Unet-Segmentation w8a8 Dragonwing Q-6690 MTP Qualcomm® Qcm6690 TFLITE 1085.57 ms 0 - 465 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 Dragonwing Q-6690 MTP Qualcomm® Qcm6690 QNN_DLC 1109.153 ms 2 - 465 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 Dragonwing RB3 Gen 2 Vision Kit Qualcomm® QCS6490 TFLITE 264.565 ms 0 - 40 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 Dragonwing RB3 Gen 2 Vision Kit Qualcomm® QCS6490 QNN_DLC 264.536 ms 4 - 10 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 121.428 ms 2 - 161 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 121.354 ms 2 - 160 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 58.258 ms 0 - 296 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 52.887 ms 2 - 299 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 33.659 ms 2 - 623 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 35.139 ms 2 - 4 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 32.272 ms 2 - 161 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 146.327 ms 2 - 161 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 SA7255P ADP Qualcomm® SA7255P TFLITE 121.428 ms 2 - 161 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 SA7255P ADP Qualcomm® SA7255P QNN_DLC 121.354 ms 2 - 160 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 SA8255 (Proxy) Qualcomm® SA8255P (Proxy) TFLITE 32.336 ms 2 - 623 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 SA8255 (Proxy) Qualcomm® SA8255P (Proxy) QNN_DLC 34.693 ms 2 - 5 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 SA8295P ADP Qualcomm® SA8295P TFLITE 63.792 ms 2 - 162 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 SA8295P ADP Qualcomm® SA8295P QNN_DLC 63.74 ms 0 - 163 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 SA8650 (Proxy) Qualcomm® SA8650P (Proxy) TFLITE 33.361 ms 2 - 404 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 SA8650 (Proxy) Qualcomm® SA8650P (Proxy) QNN_DLC 35.111 ms 2 - 4 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 SA8775P ADP Qualcomm® SA8775P TFLITE 32.272 ms 2 - 161 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 SA8775P ADP Qualcomm® SA8775P QNN_DLC 146.327 ms 2 - 161 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 25.548 ms 1 - 299 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 25.714 ms 2 - 301 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile TFLITE 21.42 ms 2 - 164 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile QNN_DLC 21.527 ms 2 - 172 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 Snapdragon 7 Gen 4 QRD Snapdragon® 7 Gen 4 Mobile TFLITE 78.435 ms 3 - 239 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 Snapdragon 7 Gen 4 QRD Snapdragon® 7 Gen 4 Mobile QNN_DLC 78.104 ms 2 - 240 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile TFLITE 15.907 ms 1 - 174 MB NPU Unet-Segmentation.tflite
Unet-Segmentation w8a8 Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile QNN_DLC 15.622 ms 2 - 179 MB NPU Unet-Segmentation.dlc
Unet-Segmentation w8a8 Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 35.757 ms 2 - 2 MB NPU Unet-Segmentation.dlc

Installation

Install the package via pip:

pip install qai-hub-models

Configure Qualcomm® AI Hub Workbench to run this model on a cloud-hosted device

Sign-in to Qualcomm® AI Hub Workbench with your Qualcomm® ID. Once signed in navigate to Account -> Settings -> API Token.

With this API token, you can configure your client to run models on the cloud hosted devices.

qai-hub configure --api_token API_TOKEN

Navigate to docs for more information.

Demo off target

The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input.

python -m qai_hub_models.models.unet_segmentation.demo

The above demo runs a reference implementation of pre-processing, model inference, and post processing.

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.unet_segmentation.demo

Run model on a cloud-hosted device

In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following:

  • Performance check on-device on a cloud-hosted device
  • Downloads compiled assets that can be deployed on-device for Android.
  • Accuracy check between PyTorch and on-device outputs.
python -m qai_hub_models.models.unet_segmentation.export

How does this work?

This export script leverages Qualcomm® AI Hub to optimize, validate, and deploy this model on-device. Lets go through each step below in detail:

Step 1: Compile model for on-device deployment

To compile a PyTorch model for on-device deployment, we first trace the model in memory using the jit.trace and then call the submit_compile_job API.

import torch

import qai_hub as hub
from qai_hub_models.models.unet_segmentation import Model

# Load the model
torch_model = Model.from_pretrained()

# Device
device = hub.Device("Samsung Galaxy S25")

# Trace model
input_shape = torch_model.get_input_spec()
sample_inputs = torch_model.sample_inputs()

pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])

# Compile model on a specific device
compile_job = hub.submit_compile_job(
    model=pt_model,
    device=device,
    input_specs=torch_model.get_input_spec(),
)

# Get target model to run on-device
target_model = compile_job.get_target_model()

Step 2: Performance profiling on cloud-hosted device

After compiling models from step 1. Models can be profiled model on-device using the target_model. Note that this scripts runs the model on a device automatically provisioned in the cloud. Once the job is submitted, you can navigate to a provided job URL to view a variety of on-device performance metrics.

profile_job = hub.submit_profile_job(
    model=target_model,
    device=device,
)

Step 3: Verify on-device accuracy

To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device.

input_data = torch_model.sample_inputs()
inference_job = hub.submit_inference_job(
    model=target_model,
    device=device,
    inputs=input_data,
)
    on_device_output = inference_job.download_output_data()

With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output.

Note: This on-device profiling and inference requires access to Qualcomm® AI Hub Workbench. Sign up for access.

Run demo on a cloud-hosted device

You can also run the demo on-device.

python -m qai_hub_models.models.unet_segmentation.demo --eval-mode on-device

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.unet_segmentation.demo -- --eval-mode on-device

Deploying compiled model to Android

The models can be deployed using multiple runtimes:

  • TensorFlow Lite (.tflite export): This tutorial provides a guide to deploy the .tflite model in an Android application.

  • QNN (.so export ): This sample app provides instructions on how to use the .so shared library in an Android application.

View on Qualcomm® AI Hub

Get more details on Unet-Segmentation's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

  • The license for the original implementation of Unet-Segmentation can be found here.

References

Community

Downloads last month
130
Inference Providers NEW
Image Segmentation
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Paper for qualcomm/Unet-Segmentation