Lesson 8: Detection on a map - territory segmentation

Approx. reading time: 25-40 minutes

In this lesson we will learn how to make territory segmentation. Proceed to a dataset with a map.
We will annotate two object types:

  • Dirt road
  • Water

The annotation toolbox is on the top bar, above the map.


The tools are the same as in the case of image annotation.

Use these tools to make 2 sets of annotations. For dirt road, please repeat the look of the following picture:


Blue shapes define road parts. Blue dashed rectangles define background.

Notice a panel on the right side. It displays all object types annotated on the map. Each object type may have several child items. “Manual” layer presents object annotations(solid line) and background(dashed line). “Detected” will appear after training and will display layer with detection results.
Each item has a “Clean” button to the right of the name.

Now make annotations for the water. Try to repeat the following picture.

Annotations Water

Also, note that you can control the visibility of each layer using checkboxes in the Layers panel.

Now press “Train” and create a new detector for the road. Then repeat the same procedure for water. Do not forget to switch “Run detector after train” to see the detection results immediately.
Also, disable “Detected objects have convex shape”. This is because the road and water (river) both have a complex geometry. For simple objects like cars or building roofs, we can keep “Detected objects have convex shape” checked.

It is possible to run multiple training sessions in parallel. Tasks are queued in the order and being processed in Atlas cloud. Depending on the queue size it may take time. But do not worry. Typically no need to wait in front of the screen. When tasks end, ATLAS sends an email notification.

After training, you could expect results similar to the picture below. All the detections made by AI are will be seen.


Also keep in mind that if you just need to run detection without training on this or other datasets, then just press “Detect” and select a proper detector.