Steps to Label
Last updated
Last updated
A labelling expert goes through the following steps to label an object. Theprovides features to aid with various steps of labelling:
,
(for a moving Obj.)
Setup point cloud visibility and density from Tool Settings (tool Footer).
Fit the entire point cloud in the viewing window using Fit to Screen (tool Header).
Jump to the centre (near the Ego vehicle) by using Jump to Origin (tool Header)
Toggle to Fullscreen if desired (tool Header).
Choose one of the two Navigation Controls to navigate in the point cloud based on your preference. (tool Header)
The tool provides multiple features & interactions that aid in object identification:
Expand and collapse the image Panel whenever needed to cross-check the object/scene-related details.
Use the help of the raycaster while navigating through the point cloud to see the respective coordinate highlighted in Camera Image(s).
On hovering on the Cameras, a red cone (Frustum) will appear in the point cloud emerging from the Ego vehicle to indicate the direction the camera is facing.
For better object visibility during Drawing/ Marking use the following features:
Identify the object either on the point cloud or on the camera sensors.
Pull up the Camera Sensors using the hotkey Shift + Q
or double-clicking the camera sensor in the reference images.
After marking the annotation, mark the annotation details in the popup that appears. Enter the Instance, Attribute Details and Relationship if any.
Once an annotation is created, the user needs to input the following:
Select the the Relation type from the dropdown.
Select the Instance it has a relationship with (multiple instances can be added).
The relationship between the two can be applied across all frames or only the current frame.
The view icon allows the user to see the list of frames the relationship exists on.
Once the relationship is saved, the visual representation of the link is displayed over the point cloud with a line connecting the two instances.
An Instance ID is an identifier assigned to each labelled object across the batch of frames.
This instance ID can be used to track an object over multiple frames.
After creating an object, the user is asked to:
Create a new instance or select + Create New.
For example, only one object will have the ID car_1 which is present in Frame 1. This car could be present in 3 Cameras and the point cloud of Frame 1. Hence each of the 4 Sensors (3 Camera, 1 LiDAR) will have an annotation with ID car_1.
The Instance Dropdown allows 2 things:
Will allow the user to create New Instances
Lists all the created instances that are missing in that Sensor (car_48 in above eg.)
Select the object.
Open the annotation details pop-up
Select + Create new'. The tool will allocate a new unique ID to it.
Save the new Instance ID. This replaces the old Instance ID in all frames for that annotation.
If the user realises that two instances are the same object, then merging of instances is required. Eg. car_2
on Frame 10 is actually car_1.
Select the annotation you want to merge into another instance. (car 2 on Frame 10).
Right-click and select Annotation Details.
Open the instance ID dropdown. This dropdown has a lists of annotation IDs missing from that current frame and the option to create a new one.
Select the appropriate existing ID that the selected one needs to merge with (car_1
).
Save and the old instance gets changed to the new selected one (car_1
becomes car_2
in Frame 10).
Merging instances is only possible when both instances have no overlapping Keyframes.
On selecting an annotation, the multiview mode will be enabled on the right-side. Hover on the multiview mode to enable the annotation to modify it from the top, side, and front views aiding accurate dimension modification.
The projection of the created cuboid in the point cloud is visible on the appropriate camera images.
Double-click on that image to enlarge it.
The 2D Instance of the same object appears under the 3D Instance created on point cloud (sharing the same ID). Eg. Truck_1 is a cuboid annotation created in the point cloud and the instance of the same object is created in Image. This will appear as Truck_1 #2 (with a 2D icon on the left of the label)
Any extra points have been labelled for an object
Or if required points have not been labelled for an object
Pull up the camera sensor Shift + Q
, locate the object, press Esc
to land in the point cloud in the same perspective as the camera sensor.
Enable the across the frames to identify and mark stationary objects.
areas of the point cloud where the object is present by marking a boundary to hide surrounding/ background points.
Hide points from bottom (ground) to top (ceiling) and vice-versa using the
Switch to a desired view to start marking an object in the top view using or 3 dimensionally using .
Click the '+' icon on the Class the object belongs to from the
All the Drawing Tools linked to the Class will appear. Select the appropriate one from the .
Enter all the mandatory attributes ( and ).
This tab is enabled if a has been defined in the recipe generator.
Learn more about the .
Instance ID can be viewed in the that appears on hover or selection of an annotation on the
Instances can be , or .
Learn how to modify annotations using .
Learn how to use .
If the Annotation created is for a moving object, use feature to mark the object across all the sequential frames without having to create it in every frame.
Use this feature when an object is required to be labeled in both point cloud and image sensors while sharing a unique .
on that object.
opens up. Choose the instance ID the same as its parent annotation created in the point cloud. Fill in the remaining details and Save.
Use the feature to locate points that have not been labelled.
Use on the class row to identify if:
Conduct auditing and mark that are observed in the
View the overall submissions of using the .