Import and run an Arm AI Portal object-detection model

Scene Detector adapters

An adapter is application code that connects the shared Android interface to a model task and runtime. Scene Detector uses one ExecuTorch adapter for supported YOLO object-detection models. The adapter selects the preprocessing and output decoder associated with the imported filename.

The model runs locally on the Android CPU. The application doesn’t upload the selected image or camera frames to a server.

Object detection in Scene Detector

Image classification assigns one or more labels to an entire image. Object detection also identifies where objects appear by returning a class, confidence score, and bounding box for each retained detection. A single street image can therefore contain separate detections for people, bicycles, motorcycles, and cars.

The confidence threshold controls which detections are displayed. Increasing the threshold removes lower-confidence boxes. Decreasing the threshold can retain more objects and more false positives.

Start with YOLOv8s because it provides the shortest path to a result. You can use another supported model after completing the default workflow. Run the commands from the project directory in the same terminal session that you used during setup.

Run YOLOv8s with ExecuTorch

Set MODEL_ID to the repository ID for the default YOLOv8s model:

    

        
        

export MODEL_ID="Arm/yolov8s-int8-xnnpack-executorch"
  

    

    

        
        

$MODEL_ID = "Arm/yolov8s-int8-xnnpack-executorch"
  

    

Download the model:

    

        
        

export MODEL_FILE="$(.hf-venv/bin/python download_model.py \
  --repo-id "$MODEL_ID" \
  --print-path)"

printf 'Model file: %s\n' "$MODEL_FILE"
  

    

    

        
        

$MODEL_FILE = .\.hf-venv\Scripts\python.exe download_model.py `
  --repo-id "$MODEL_ID" `
  --print-path

Write-Output "Model file: $MODEL_FILE"
  

    

The downloader selects yolov8s_raspberry_executorch_optimized.pte, the filename recognized by Scene Detector. Keep the filename unchanged because the application uses it to select the YOLOv8 preprocessing and decoder.

Copy the model to the phone’s Downloads directory:

    

        
        
adb push "$MODEL_FILE" /sdcard/Download/

    

Copy the sample image to Android

Download the street-scene image from Wikimedia Commons and save it as street-scene.jpg. The image contains pedestrians, bicycles, a motorcycle, and cars, which gives the detector several object classes to find.

The image is by Jonas Kimmich and is licensed under CC BY-SA 4.0 .

Download the sample image:

    

        
        

curl --fail --location --output street-scene.jpg "https://upload.wikimedia.org/wikipedia/commons/thumb/7/71/DSC_6799-_a_man_riding_a_motorcycle_down_a_street_next_to_a_parked_car.jpg/1280px-DSC_6799-_a_man_riding_a_motorcycle_down_a_street_next_to_a_parked_car.jpg"
  

    

    

        
        

curl.exe --fail --location --output street-scene.jpg "https://upload.wikimedia.org/wikipedia/commons/thumb/7/71/DSC_6799-_a_man_riding_a_motorcycle_down_a_street_next_to_a_parked_car.jpg/1280px-DSC_6799-_a_man_riding_a_motorcycle_down_a_street_next_to_a_parked_car.jpg"
  

    

Copy the image to the phone:

    

        
        
adb push street-scene.jpg /sdcard/Download/

    

To use a different JPEG or PNG image, replace street-scene.jpg in the adb push command with its path. You can also skip the copy command and select an image already stored on the phone or emulator.

Run object detection

In Scene Detector:

1. Select Add or change model and choose the .pte file from Downloads.
2. Select Choose saved image and choose street-scene.jpg from Downloads.
3. Keep the default 75% confidence threshold for the first run.
4. Select Detect objects.

The application runs detection once and displays the annotated image, detected objects, confidence scores, and processing time.

The output is similar to:

Scene Detector analyzing the sample street image with YOLOv8s

Increase the confidence threshold to remove more lower-scoring boxes. Reduce it if the model doesn’t display expected objects.

Use the live camera

On the phone, select Start live camera and allow camera access. The application uses the most recent camera frame and drops queued frames while inference is running. Dropping queued frames prevents a slow detector from building an increasingly delayed frame queue.

The live-camera view draws each retained label, confidence score, and bounding box over the preview.

Camera behavior on an emulator depends on its configured front and back camera sources. Use the saved-image workflow when the emulator doesn’t expose a useful camera feed.

Use another supported model

YOLOv8s keeps the main workflow predictable. Scene Detector also supports the following Arm AI Portal model packages:

To use one of these models, set MODEL_ID to a repository ID from the table and repeat the download and import workflow.

What you’ve accomplished and what’s next

You’ve imported an optimized ExecuTorch detector, analyzed a saved image locally on Android, and used the same detector with a live camera feed.

Next, you’ll understand how Scene Detector works.