Skip to content

Azure AI Custom Vision - Object Detection

Problem

You need to automatically identify and locate specific objects in an image—for example, detecting different items on a grocery store checkout belt.

Solution with Azure

Use Azure AI Custom Vision to build, train, and deploy an object detection model that identifies classes of objects and their positions in an image.

Componenti richiesti

  • Azure AI Custom Vision Training resource - To upload, label, and train custom object detection models.
  • Azure AI Custom Vision Prediction resource - To generate object detection predictions using the trained model.
  • Custom Vision portal - A graphical web interface for project creation, labeling, training, and model management.
  • Custom Vision SDK or REST API - To programmatically interact with training and prediction services.

Architettura/Sviluppo

Training Workflow

  1. Create a Custom Vision project via portal or SDK.
  2. Upload training images.
  3. Label each image with tags and bounding boxes (either manually, via smart labeler, or external tools).
  4. Train the model and publish it to a prediction resource.
  5. Evaluate model performance using provided evaluation tools in the portal.

Labeling Details

  • Tags are associated with specific bounding boxes, not entire images.
  • Bounding boxes are defined with four values:
  • left, top, width, height (proportional to the image dimensions).

Prediction Workflow

Using Python SDK:

from msrest.authentication import ApiKeyCredentials
from azure.cognitiveservices.vision.customvision.prediction import CustomVisionPredictionClient

credentials = ApiKeyCredentials(in_headers={"Prediction-key": "<YOUR_PREDICTION_RESOURCE_KEY>"})
prediction_client = CustomVisionPredictionClient(endpoint="<YOUR_PREDICTION_RESOURCE_ENDPOINT>", credentials=credentials)

image_data = open("<PATH_TO_IMAGE_FILE>", "rb").read()
results = prediction_client.detect_image("<YOUR_PROJECT_ID>", "<YOUR_PUBLISHED_MODEL_NAME>", image_data)

for prediction in results.predictions:
    if prediction.probability > 0.5:
        print(f"{prediction.tag_name} ({prediction.probability})")
        print(f" Left:{prediction.bounding_box.left}, Top:{prediction.bounding_box.top}, Height:{prediction.bounding_box.height}, Width:{prediction.bounding_box.width}")

Best practice / Considerazioni

  • Separate training and prediction resources for regional optimization and cost efficiency.
  • Use smart labeler after initial training to speed up further labeling.
  • If using external labeling tools, ensure the format matches Custom Vision API requirements (normalized bounding box values).
  • Use appropriate SDKs for your development language (e.g., azure-cognitiveservices-vision-customvision for Python).

Domande simulate d'esame (basate solo sul contenuto fornito)

  1. What is required to use Azure AI Custom Vision for object detection?

  2. A training resource and a prediction resource in Azure.

  3. What distinguishes object detection from image classification in Custom Vision?

  4. Object detection uses bounding boxes with tags; image classification uses tags for the entire image.

  5. How are bounding boxes defined in Custom Vision labeling?

  6. By proportional values for left, top, width, and height relative to the image.

  7. What is the purpose of the Custom Vision portal?

  8. To create, label, train, publish, and evaluate object detection models.

  9. Which SDK package is used for object detection prediction in Python?

  10. azure-cognitiveservices-vision-customvision

  11. Can training and prediction resources be located in different regions?

  12. Yes, to support flexible deployment scenarios.