Previously we discussed different types of segmentation and popular machine learning use cases where segmentation is critical for the outcomes. Semantic segmentation is often a very time consuming task and can eat into timelines if you don’t have access to good annotation tools that facilitate simplification and improve accuracy of the annotations. Today, we’ll explore GT Studio features that simplify segmentation tasks for annotators.
GT Studio is Playment’s proprietary, web-based data annotation platform that allows ML teams to leverage state-of-the art labeling tools and project management software on a single integrated platform.
Here are some features that make segmentation on GT Studio simpler and more accurate.
1. AI-Assisted Labeling
Instead of marking numerous points along the boundary of an object, the AI-assisted labeling feature lets you mark the four farthest corners of the object to automatically create a semantic mask around the object of interest. The points must be marked in an order, refer image below.
The semantic masks are accurate and also saves time when annotators are segmenting multiple objects on a single frame across longer sequences.
Here’s an example of AI-Assitance being used to mark a car with just four clicks for an autonomous driving use case.
2. Flood Fill
Flood fill is a selection technique that allows users to select a patch with similar colored pixels with just one click. The user also has the control to increase or decrease the selected patch by clicking and dragging the cursor in different directions. Lastly, press Enter and the highlighted area gets filled with the color of the selected instance.
This feature is especially useful in segmenting objects covering larger sections of an image, for example, sky, vegetation, road, etc.
3. Smoothen Pixels
When an annotator creates a semantic mask, there are multiple points marked on the object boundary. With a simple toggle on/off button, you can create a more accurate boundaries or smoothen pixels around the object. This feature multiplies the points, creating more sides to the semantic mask/polygon, providing a more refined segment of the object.
4. Reference image support
Annotators lose a lot of time switching between reference images on multiple tabs while labeling. With the new reference image support feature, you can now easily pull up references while segmenting primary images in the tool. The references help you annotate or segment images without leaving the tool interface.
5. Draw on top
This feature lets you override segments to create new polygons or semantic masks just like you would using layers. It makes it very simple to draw over segments while segmenting new sections of an image.
6. Show/hide Unmarked Areas
To avoid missing segments in an image, we’ve added this feature where the unmarked areas begin to pop on the screen to help guide you to the sections that are to be segmented. This ensures that every pixel in the image is segmented before the annotator moves on to the next.
7. Segmentation with added attributes
Additional attributes help make the segmented data richer and more useful for ML models. That’s why we’ve introduced attributes for segmentation. You can add unlimited attributes to the instances and segments on an image.
For example, in addition to marking an instance of Person 1, we can also classify the the occlusion (0 - 50%; 50%-100%), state (standing, sitting, walking, riding, others) and various other properties of the instances.
8. View by Groups
Reviewing segmentation tasks aren’t usually simple. But with the View by Group feature you can check if segments and their corresponding instances and attributes are marked correctly in a glimpse by viewing them by class/instance groups and attribute groups.
9. Progress Bar
The progress bar provides information about the percentage of the image that has been segmented. This feature further helps avoid unmarked or non-segmented sections in the images.
10. Doodles & Comments
The doodle feature allows a reviewer to draw a doodle over a mistake and provide comments to the annotator so they can improve the segmentation outputs right away. This feature facilitates instant feedback cycles and ensure the throughputs are higher and the output is more accurate.
For quality control tasks, the reviewer or quality analyst can mark doodles and also additionally classify errors as critical or non-critical along with providing more information about the errors. The classification of errors provide insights into the execution process and how it can be improved in the future.
Try GT Studio for free today...
Anyone can sign up and start using GT Studio. We have an excellent team to support you through your journey while exploring our platform.
Feel free to reach us at email@example.com