Is there a way to upload images with annotations (labeled images) to custom vision?

Question

I have hundreds of labeled images and do not want to redo that work in the custom vision labeling tool. Is there a way to upload labeled images to custom vision? Or to Azure ML or Azure ML Studio? Does any vision services in Azure provide for uploading annotated images? Thanks

Answer 1

I built a proof of concept using a package that I developed called PyLabel to upload annotations to Azure Custom Vision. You can see it here https://github.com/pylabel-project/samples/blob/main/pylabel2azure_custom_vision.ipynb.

PyLabel can read annotations from COCO, YOLO, or VOC format into a dataframe. Once they are in the data frame you can loop through the dataframe of annotations and use the Custom Vision APIs to upload the images and annotations.

The annotation format used by Custom Vision similar to the YOLO format because they both used normalized coordinated between 0-1.

Here is a snippet of the code from the notebook mentioned above:

#Iterate the rows for each image in the dataframe
for img_filename, img_df in dataset.df.groupby('img_filename'):
    img_path = str(PurePath(dataset.path_to_annotations, str(img_df.iloc[0].img_folder), img_filename))
    assert exists(img_path), f"File does not exist: {img_path}"

    #Create a region object for each bounding box in the dataset 
    regions = []
    for index, row in img_df.iterrows():

        #Normalize the boundings box coordinates between 0 and 1
        x = Decimal(row.ann_bbox_xmin / row.img_width).min(1)
        y = Decimal(row.ann_bbox_ymin / row.img_height).min(1)
        w = Decimal(row.ann_bbox_width / row.img_width).min(1-x)
        h = Decimal(row.ann_bbox_height / row.img_height).min(1-y)
        
        regions.append(Region(
                tag_id=tags[row.cat_name].id, 
                left=x,
                top=y,
                width=w,
                height=h
            )
        )

    #Create an object with the image and all of the annotations for that image
    with open(img_path, mode="rb") as image_contents:
        image_and_annotations = [ImageFileCreateEntry(name=img_filename, contents=image_contents.read(), regions=regions)]

    #Upload the image and all annnotations for that image
    upload_result = trainer.create_images_from_files(
            project.id, 
            ImageFileCreateBatch(images=image_and_annotations)
        )
    
    #If upload is not successful, print details about that image for debugging 
    if not upload_result.is_batch_successful:
        print("Image upload failed.")
        for image in upload_result.images:
            print(img_path)
            print("Image status: ", image.status)
            print(regions)

#This will take a few minutes 
print("Upload complete")