Contents x
    Model Assisted Annotations
    • 15 May 2023
    • Dark
      Light
    • PDF
    Contents

    Model Assisted Annotations

    • Updated On 15 May 2023
    • Dark
      Light
    • PDF
    Article Summary

    ML/AI assisted annotations helps accelerate the human touch, reducing human time, effort, and cognition invested by up to 90%!

    With Dataloop, you can integrate existing models or run our pre-trained models to automatically pre-label complete datasets.
    Automation pipelines are the most efficient time and cost saving element enabling massive data reduction via smart pre and post processing flows.

    Annotation Studio Level Processing

    By using one of Dataloop's pre-configured models, or your own custom model, annotators can activate a prediction function on an item according to need.

    Automation models are available in the Image and Video annotation studios only

    Pre-Configured Models

    When enabled from a task recipe, an annotator can select the model/function to run on the item currently active in the annotation studio. 

    To activate a pre-configured annotation model:

    1. Make sure that automations are enabled from recipe instructions
      image.png
    2. Enter an annotation studio
    3. From the left-side tool-selection panel, select the Automation tool
    4. Select the model to run from the dropdown below
      image.png

    Run Your Automation Model

    By installing your own model and deploying it to the respective UI-Slot, your model can be used by annotators in the same way as Dataloops' default models. 

    For more information, follow this tutorial to add your model as a FaaS package, and deploy it to a UI slot

    Filtering Model Annotations By Confidence

    When working with automated annotations in the annotation studio, Dataloop provides a filter that allows filtering annotations by their confidence score. The filter works with a minimum and maximum range, so you can filter out the lowest score annotations and focus on the highest ones, or filter out the high score ones to be able to select and delete all the lower-score annotations.

    1. In the annotation studio's right-side annotations panel, locate the Group-Mode button and switch it to "By Creator" - Annotations will be grouped by its creator and not by label, hence group by models.
    2. On every model's group. locate the filter dropdown to see the confidence filter range selection. Adjust the Min and Max values, and annotations will filter in/out of the studio.


    Pre-Processing - Model Annotations

    Dataloop enables you to preprocess all the data in advance as per your project needs (for example as part of Pipeline processes, or to simply avoid having annotators manually running it from the annotation studio).

    Before you start:

    1. Install your model and run it as a service - follow this tutorial to learn how
    2. Deploy your model service to relevant UI slots - read here to learn how  

    Run Model On Data (Dataset Browser)

    From the Dataset browser
    1. Filter the items you intend to preprocess (read the Data Filtering tutorial)
    2. Ensure your model service is deployed to the Dataset-browser UI-Slot
    3. Select the Automation tool from the top actions bar, and select your model from the dropdown

    Manually Trigger Data To Model Service

    1. From the Dataset browser, select 1 or more items
    2. From the right-side panel, select the relevant Function and click on EXECUTE

    Preprocess With Data Pipelines

    For more information about working with Pipelines, read the respective documentation section.

    1. From the left navigation menu, select the Pipelines menu item
    2. Start a new Pipeline by pressing on "CREATE PIPELINE", and either start a new empty one or select from an available template
    3. Drag & drop a FaaS node onto the pipeline canvas, and from node settings select your model function from the list, to et this node to run your model prediction function
    4. Send data to the Pipeline by triggering it or by connecting a dataset node before the FaaS one, and your items will go through the Pipeline and processed - model annotations will be saved on the item
    5. Finish your pipeline composition to achieve and other preprocessing steps and technical/business goals

    Upload Model Annotations

    Using Dataloops' SDK, you can upload annotations containing model information and annotation confidence score. The process is easy and straightforward, using the Builder for creating annotations in the Dataloop format.

    Here's an example of creating and uploading a bounding-box annotation using the builder, with model information and confidence score:

    builder.add(
       annotation_definition=dl.Box(
          top=startY,
          left=startX,
          right=endX,
          bottom=endY,
          label='person'
       ), 
       model_info={
           'name': 'Caffe',
           'confidence': confidence
       }
)   


    What's Next