When annotating millions of images for deep learning projects, it's easy to lose track of the details. Comments help your team communicate asynchronously and maintain a higher dataset quality.
The comment tool is very straightforward - select a region, leave a comment, get notified of a reply. We developed this tool for a number of reasons:
• Questions create friction. Rather than interrupting one's workflow to ask a question about a specific image, an annotator can leave a comment, move on, and await for a reviewer or another team member to respond.
• Comments are great for quality control. Reviewers can leave comments to explain why an image was rejected. During review, pressing R to reject an image won't cycle you automatically to the next to encourage you to add a comment (hotkey: C) for the annotation author so they can learn how to improve.
• Comments help teams insert instructions within images - try this with the first 5 images of a dataset.
• They are essential to query and store the opinion of experts on an image annotation. This last point is particularly important for medical use cases.
The information found in comments is not used when training models. In fact, most of the time you will want to resolve comments before sharing or exporting a dataset. Nonetheless they can be useful when you re-investigate ambiguous image that don't perform well in a test set - perhaps the ground truth itself has some contesting opinions.