Work with empty annotations¶

In this article, you will learn how to use empty annotations and distinguish them from no annotations.

Concept of "empty annotation"¶

Glossary

Empty annotations are used when a sample doesn't contain any of the classes present in the project. Assigning an empty annotation indicates that none of the existing classes apply to the sample.

Project types that support empty annotations:

Multilabel classification
Object detection
Instance segmentation
Semantic segmentation
Multiview instance segmentation

Example

Imagine you are working on a project to identify and classify animals in images. You have three classes: dog, cat, and bird. If you come across an image that doesn't contain any of these animals (for example, an image of a car), you can assign an empty annotation to that sample, indicating that none of the existing classes apply to this image.

Concept of "no annotation"¶

A sample with no annotation is a sample that hasn't been annotated, or its annotation has been deleted by the user. It represents unprocessed or overlooked samples.

"No annotation" vs "empty annotation"¶

The choice between "no annotations" and an "empty annotation" depends on your task.

	No annotation	Empty annotation
Advantages	Simplicity in training data, which might be beneficial for tasks where minimal guidance is sufficient. Avoids the potential introduction of noise or misleading information from empty annotations.	Provides a structured framework for future annotations, which can be valuable for tasks where additional details might be introduced later. Can be useful for tasks requiring a consistent format, even if the annotations are currently empty.
Disadvantages	Lack of explicit guidance may result in the model struggling to learn patterns or understand the context, especially in tasks requiring detailed information.	May introduce unnecessary complexity and potentially confuse the model if the structure of the empty annotation is not clearly defined.

Considerations:

If the task requires clear and immediate guidance, having "no annotations" may be preferable.
If the task involves evolving annotations or future updates, providing an "empty annotation" framework can be beneficial.

In the user interface¶

Empty annotations are visualized with a crossed-out black diamond.
When written out, they are called "empty annotation".

Assign an empty annotation¶

In the single view:
1. On the bottom toolbar, click Annotate as empty.
2. Click Submit.
In the thumbnails or list view:
1. Select one or several samples.
  
  Note
  
  If you select samples and then go to another page, these samples will no longer be selected.
2. On the bottom toolbar, click Annotate as empty.

Filtering and datasets¶

Using filters: In the Filter panel, you can filter samples with annotations, empty annotations, and no annotations. Annotations also include empty annotations.
Class distribution chart: In the class distribution chart, you can view and filter samples with empty annotations and no annotations.
Creating datasets: When creating datasets, samples with empty annotations are added to the datasets.

Key takeaways¶

Use empty annotations for samples that don't contain any of the classes in your project.
No annotations represent unprocessed or overlooked samples in your project.
Use filters and charts to view and analyze samples with empty annotations and no annotations in your dataset.
Add samples with empty annotations to your datasets acknowledging that these samples don't contain any of the project classes.