Point cloud Segmentation

How to perform semantic and instance segmentation annotation for point clouds on Mindkosh.

Create a new segmentation object

To create a new segmentation object on a pointcloud, you can use the Lasso tool as described below.

1. Choose a label for the object from the left sidebar.

2. Select the Draw segmentation tool from the left sidebar.

3. Hold the Right mouse button, and draw a rough outline around the points you want to select.

4. A segmentation with the selected points is drawn.

Editing an existing segmentation object

1. Select the object.

   1. You can select the object by clicking on the object card on the right.

   2. Or you can also select an object by holding G, and clickin on any of its points. Make sure you are in the Move mode to make this work. You can activate the Move mode by clicking the Move tool button from the left sidebar.

2. Hold the Right mouse button, and draw a rough outline around the points you want to add to the object , or you want to remove from the object.

3. The selected points will be highlighted in red. If you want to cancel the operation, press the ESC key

4. With the points selected, the Add points and Remove points buttons will appear inside the object card in the right sidebar.

5. Click on one of the buttons to update the segmentation.

Create segmentation from cuboids

You can also create segmentation objects from cuboids. To do this, first draw a cuboid around the points, then click on the Create Seg Object button in the object card from the right. This will add all points within the cuboid to a new Segmentation object.

This can be very handy if your objects are in roughly cuboid shapes, or if you can use 1-click annotation to quickly label objects with cuboids.

Hide and filter points

When creating annotations for segmentation, it can be helpful to isolate and identify points assigned a certain label.

Filter points by class

To entirely hide the points of a certain class, click on the filter icon for that class from the left sidebar. This can be very helpful to annotate to adjacent classes. Once you have labeled an object, you can hide it and proceed with labeling the neighbouring objects.

To aid segmentation of point clouds, we also provide an All labeled filter at the end of the label list. This will hide or show all labeled points. You can use this quickly see if any points were missed while labeling, while also making it easier to label by reducing clutter.