BiseNet: Optimized for Mobile Deployment
Segment images or video by class in real-time on device
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 model is an implementation of BiseNet found here.
This repository provides scripts to run BiseNet 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: best_dice_loss_miou_0.655.pth
- Inference latency: RealTime
- Input resolution: 720x960
- Number of parameters: 12.0M
- Model size (float): 45.7 MB
| Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |
|---|---|---|---|---|---|---|---|---|
| BiseNet | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 133.376 ms | 32 - 69 MB | NPU | BiseNet.tflite |
| BiseNet | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 111.86 ms | 3 - 75 MB | NPU | BiseNet.dlc |
| BiseNet | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 55.56 ms | 32 - 79 MB | NPU | BiseNet.tflite |
| BiseNet | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN_DLC | 63.236 ms | 8 - 60 MB | NPU | BiseNet.dlc |
| BiseNet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 46.549 ms | 32 - 49 MB | NPU | BiseNet.tflite |
| BiseNet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 32.723 ms | 8 - 26 MB | NPU | BiseNet.dlc |
| BiseNet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 33.304 ms | 63 - 122 MB | NPU | BiseNet.onnx.zip |
| BiseNet | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 56.389 ms | 32 - 69 MB | NPU | BiseNet.tflite |
| BiseNet | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 41.608 ms | 2 - 73 MB | NPU | BiseNet.dlc |
| BiseNet | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 133.376 ms | 32 - 69 MB | NPU | BiseNet.tflite |
| BiseNet | float | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 111.86 ms | 3 - 75 MB | NPU | BiseNet.dlc |
| BiseNet | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 46.029 ms | 20 - 38 MB | NPU | BiseNet.tflite |
| BiseNet | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN_DLC | 32.78 ms | 8 - 28 MB | NPU | BiseNet.dlc |
| BiseNet | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 61.008 ms | 32 - 74 MB | NPU | BiseNet.tflite |
| BiseNet | float | SA8295P ADP | Qualcomm® SA8295P | QNN_DLC | 47.183 ms | 6 - 56 MB | NPU | BiseNet.dlc |
| BiseNet | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 46.075 ms | 20 - 38 MB | NPU | BiseNet.tflite |
| BiseNet | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN_DLC | 33.258 ms | 8 - 31 MB | NPU | BiseNet.dlc |
| BiseNet | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 56.389 ms | 32 - 69 MB | NPU | BiseNet.tflite |
| BiseNet | float | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 41.608 ms | 2 - 73 MB | NPU | BiseNet.dlc |
| BiseNet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 31.893 ms | 31 - 76 MB | NPU | BiseNet.tflite |
| BiseNet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 24.512 ms | 8 - 66 MB | NPU | BiseNet.dlc |
| BiseNet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 26.227 ms | 73 - 129 MB | NPU | BiseNet.onnx.zip |
| BiseNet | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 26.62 ms | 30 - 73 MB | NPU | BiseNet.tflite |
| BiseNet | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNN_DLC | 20.068 ms | 8 - 77 MB | NPU | BiseNet.dlc |
| BiseNet | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 19.324 ms | 65 - 121 MB | NPU | BiseNet.onnx.zip |
| BiseNet | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 22.725 ms | 31 - 73 MB | NPU | BiseNet.tflite |
| BiseNet | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNN_DLC | 17.424 ms | 8 - 79 MB | NPU | BiseNet.dlc |
| BiseNet | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 15.303 ms | 73 - 136 MB | NPU | BiseNet.onnx.zip |
| BiseNet | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 31.161 ms | 8 - 8 MB | NPU | BiseNet.dlc |
| BiseNet | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 31.68 ms | 66 - 66 MB | NPU | BiseNet.onnx.zip |
| BiseNet | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 24.46 ms | 8 - 46 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 21.383 ms | 2 - 55 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 16.485 ms | 8 - 70 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN_DLC | 18.298 ms | 2 - 69 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 15.023 ms | 8 - 19 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 10.391 ms | 2 - 20 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 10.428 ms | 8 - 55 MB | NPU | BiseNet.onnx.zip |
| BiseNet | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 14.981 ms | 8 - 46 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 11.001 ms | 2 - 53 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | TFLITE | 54.89 ms | 8 - 68 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | QNN_DLC | 55.786 ms | 2 - 87 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | ONNX | 220.532 ms | 206 - 219 MB | CPU | BiseNet.onnx.zip |
| BiseNet | w8a8 | RB5 (Proxy) | Qualcomm® QCS8250 (Proxy) | TFLITE | 287.541 ms | 8 - 11 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | RB5 (Proxy) | Qualcomm® QCS8250 (Proxy) | ONNX | 194.766 ms | 198 - 224 MB | CPU | BiseNet.onnx.zip |
| BiseNet | w8a8 | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 24.46 ms | 8 - 46 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 21.383 ms | 2 - 55 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 14.51 ms | 8 - 24 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN_DLC | 10.358 ms | 2 - 23 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 17.559 ms | 8 - 50 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | SA8295P ADP | Qualcomm® SA8295P | QNN_DLC | 13.553 ms | 2 - 56 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 14.662 ms | 7 - 22 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN_DLC | 10.396 ms | 2 - 21 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 14.981 ms | 8 - 46 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 11.001 ms | 2 - 53 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 10.798 ms | 6 - 64 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 7.441 ms | 2 - 75 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 7.523 ms | 18 - 88 MB | NPU | BiseNet.onnx.zip |
| BiseNet | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 8.284 ms | 7 - 50 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNN_DLC | 5.769 ms | 2 - 61 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 5.705 ms | 16 - 74 MB | NPU | BiseNet.onnx.zip |
| BiseNet | w8a8 | Snapdragon 7 Gen 4 QRD | Snapdragon® 7 Gen 4 Mobile | TFLITE | 20.504 ms | 0 - 53 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | Snapdragon 7 Gen 4 QRD | Snapdragon® 7 Gen 4 Mobile | QNN_DLC | 16.35 ms | 2 - 72 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | Snapdragon 7 Gen 4 QRD | Snapdragon® 7 Gen 4 Mobile | ONNX | 213.382 ms | 212 - 227 MB | CPU | BiseNet.onnx.zip |
| BiseNet | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 7.193 ms | 6 - 48 MB | NPU | BiseNet.tflite |
| BiseNet | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNN_DLC | 4.903 ms | 2 - 65 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 4.539 ms | 18 - 85 MB | NPU | BiseNet.onnx.zip |
| BiseNet | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 10.918 ms | 11 - 11 MB | NPU | BiseNet.dlc |
| BiseNet | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 10.927 ms | 19 - 19 MB | NPU | BiseNet.onnx.zip |
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.bisenet.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.bisenet.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.bisenet.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.bisenet 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.bisenet.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.bisenet.demo -- --eval-mode on-device
Deploying compiled model to Android
The models can be deployed using multiple runtimes:
TensorFlow Lite (
.tfliteexport): This tutorial provides a guide to deploy the .tflite model in an Android application.QNN (
.soexport ): This sample app provides instructions on how to use the.soshared library in an Android application.
View on Qualcomm® AI Hub
Get more details on BiseNet's performance across various devices here. Explore all available models on Qualcomm® AI Hub
License
- The license for the original implementation of BiseNet can be found [here](This model's original implementation does not provide a LICENSE.).
- The license for the compiled assets for on-device deployment can be found here
References
- BiSeNet Bilateral Segmentation Network for Real-time Semantic Segmentation
- Source Model Implementation
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.
- Downloads last month
- 216