How do I upload annotated dataset from Roboflow to Custom Vision Project?

Hello Maheen Rasool,

Yes, you can upload an annotated dataset from Roboflow to a Custom Vision project in Azure, but the process is not directly seamless because Azure Custom Vision expects a specific format. Here's a step-by-step explanation on how to make it work:

Step-by-Step Guide: Upload Roboflow Dataset to Azure Custom Vision

Step1: Download Dataset from Roboflow

·       In your Roboflow project:

o   Click Export Dataset.

o   Choose the format: Select "VOC XML" (Pascal VOC format), which is supported by Azure Custom Vision.

o   Choose image size (original size is preferred).

o   Download the ZIP file.

Step2: Understand Azure Custom Vision Requirements

Azure Custom Vision accepts:

·       Images (.jpg, .png, etc.)

·       Annotation file: For object detection, it accepts Pascal VOC (XML files with bounding boxes).

·       All images and XML files should be uploaded via the Custom Vision Portal or Python SDK / REST API.

Step3: Extract and Organize the Dataset

After unzipping the dataset:

·       You’ll get two folders:

o   images/ – contains image files

o   annotations/ – contains Pascal VOC XML files

Make sure:

·       Each image has a corresponding .xml annotation file with the same name (e.g., car_001.jpg ↔ car_001.xml).

·       All files are clean and consistent.

Step4: Create a Custom Vision Project

1.      Go to the Azure Custom Vision Portal.

2.      Click New Project.

3.      Choose:

o   Project Type: Object Detection

o   Domain: General (or Retail, etc., based on your use-case)

o   Resource Group & Training Resource: As configured in Azure.

Step5: Upload Data to Custom Vision

Option1: Using Azure Custom Vision Portal (GUI)

·       Not ideal for large datasets with annotations.

·       You can upload images and manually tag them – time-consuming.

Option2: Use the Custom Vision Python SDK

Here’s a basic script to upload images + Pascal VOC annotations:

from azure.cognitiveservices.vision.customvision.training import CustomVisionTrainingClient
from msrest.authentication import ApiKeyCredentials
import os
# Setup
ENDPOINT = "https://<your-customvision-endpoint>.cognitiveservices.azure.com/"
TRAINING_KEY = "<your-training-key>"
PROJECT_ID = "<your-project-id>"
credentials = ApiKeyCredentials(in_headers={"Training-key": TRAINING_KEY})
trainer = CustomVisionTrainingClient(ENDPOINT, credentials)
# Upload dataset
images_folder = "images"
annotations_folder = "annotations"
for image_filename in os.listdir(images_folder):
    if image_filename.endswith(".jpg"):
        image_path = os.path.join(images_folder, image_filename)
        annotation_path = os.path.join(annotations_folder, image_filename.replace(".jpg", ".xml"))
        
        with open(image_path, "rb") as image_contents:
            regions = trainer.create_regions_from_voc(annotation_path)
            trainer.create_images_from_data(
                project_id=PROJECT_ID,
                images=[{
                    "name": image_filename,
                    "contents": image_contents,
                    "regions": regions
                }]
            )

Step6: Train Your Model

Once images are uploaded with bounding boxes:

·       Click Train in the portal or use:

iteration = trainer.train_project(PROJECT_ID)

Attaching links below for your reference

https://docs.roboflow.com/datasets/dataset-versions/exporting-data

https://customvision.ai/projects

Hope this helps, please do up-vote and accept if this helps in resolving your query.

Thank you!