The quality of your Training Data is defined by the quality of annotations and labels. Your annotators can label your training data with the tools you (or your data labeling software vendor) provide them.
The right instance type will help you label easily and fast. It enables your crew to label more images in a very short time.
Diffgram hosts a wide range of instance tools that make annotating or labeling images very simple. The following are all the various types of instances available to annotate your images on Diffgram.
Various image annotation instance types to annotate your image training data
1. Bounding Box:
A bounding box is the most basic and very widely used instance type. This tool allows you to create and use boxes to label your data. Bounding boxes are helpful to provide high level information about your image and the information it represents. Bounding boxes are not worried about the tight pixelization and are used to encase labels of information in rectangular or square boxes.
To start using bounding boxes on Diffgram, make sure that you are in the drawing mode and have selected the label you wish to annotate your image with. Point and click where you want to start the box; drag and click to end the box. You now have a bounding box. To adjust this, you can click on the drawing mode to change it to editing.
The Polygon tool is basically like a pencil tool. It allows you to create very detailed annotations. You can draw it to add very specific boundaries for your labels. Complex Polygon annotations are used to improve the accuracy of your labels. These labels can be very tight to their borders and edges and are often used for almost every non-conforming instance types.
For example, you cannot use a bounding box to label a bacteria or a complex shaped country boundary or a vehicle engine etc.,
To get started, select the “polygon” instance type and also with the appropriate label; in the case of the video below, it’s a car.
You click to start and click along the way of your borders. As you will see, the polygon tool adds a straight line between two clicks. If you wish to annotate closer or irregular edges faster, you can press and hold the shift key to enter into the turbo mode.
Turbo Mode on Diffgram:
Turbo mode will make the process faster. I press the shift key here. Hold it till here. Then release these points and then hold it back. My label is now done much faster. You can also use the hotkey shift+t to add or remove the transparency. This gives you a good idea on if the labeling is done right.
Fixing/Editing A Poorly Annotated Polygon:
Not 100% of your annotations are perfect in the first try. While creating a polygon annotation, you might have missed a border or have drawn out of line. To edit your polygon annotation, you can go to edit mode, zoom in and double click on the points that are out of place.The tool will automatically delete those instances.
If there’s a need to add a new point anywhere in between two points, simply click once and it will automatically add a new point. You can also drag these points to snap them together in their intended places.
This is particularly useful when you are re-labeling your previously complexly annotated training data.
Points are the most simplest instance types. They are also called as landmark or dot annotations in various platforms. They are mostly useful in instances where you need to just put tiny labels and do not need bounding boxes or polygons. For example, you can label people in a crowd or very small objects like pebbles or stars or bulbs. Basically anything that does not need a polygon or encasing.
It is simple to use a dot annotation. You just point and click and you’re done.. In the video below, we try to label people in a given crowd to see if they are a man or a woman.
4. Fixed Lines:
Fixed line annotation, as the name suggests, this instance type gives you a fixed straight line for annotating your data.
5. 2D Cuboids:
2D Cuboids can be used to create annotation instances that go far beyond a simple bound box. A 2D cuboid can try and wrap through a 3D area in a 2D Image. This gives you enough breadth to label images that is not quite easy with a simple bounding box. You can see it in action below.
There are plenty more to be explored
Diffgram is constantly evolving and will keep trying to add anything that makes the life of your annotator easier. Apart from the above, we have plenty other instance tools such as Ellipse & Circle, Curve Quadratic etc.,
Try Diffgram Online or intall it locally.