Object Detection & Tracking

Guide for Instance Annotation Mode.

To detect objects from LiDAR Fusion Data or track them in a Scene, you need to choose Instance (Alt+1) in the Mode Switch.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

📘

Recommended Workflow for Object Detection & Tracking

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

1. Preparation

Before starting the annotation, you may need to adjust the view and display settings first.

The point cloud's focal range is often not at the coordinate origin. You can continuously press the hotkey Y multiple times to reposition it.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Clicking on the eye icon on the image data will highlight the camera's field of view on the Point Cloud Data.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Press the hotkey :; to show/hide the Linkage Pointer.

When you move the mouse over images, a line will appear in the point cloud canvas, indicating the direction of the pointer. Conversely, a linkage point will appear on the image when you move the mouse over the point cloud data.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

At the bottom of the Toolbar, you can adjust the display settings, including the background color, ground of the point cloud data, point and object display, and other utilities to help identify objects more accurately.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

2. Detection in 3D

Select tools from the Toolbar to detect objects on the Point Cloud Data. You can modify the results in Orthographic Views.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Cuboid

Select the 3D Box tool (F) to create a cuboid. You can click ... below the icon or hit the U key to switch between two methods.

  • AI-assisted Boxing: Click on the point cloud data to anchor the starting and ending points. AI will automatically detect the object and create a cuboid around it.
  • Manual Boxing: Anchor three points to generate a cuboid.

The orientation (head) of the cuboid is determined by the starting and ending points. You can hit the hotkey C or Shift+C to rotate or flip it.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

On the right are three views of the target object, allowing independent scaling and translation. You can drag the box to rescale the result or adjust its position using buttons/shortcuts.

HotkeyViewFunction
Z/XOverheadRotate the cuboid clockwise / counterclockwise
W/S/A/DSideMove the cuboid up/down/backward/forward
Q/ERearMove the cuboid left/right
annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

To adjust the cuboid position, you can also click on Translate (G) in the toolbar. Move the cuboid by dragging it or translating it along the coordinate axes.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Polygon

Select the Polygon tool from the toolbar. Click on the canvas to add points that will be connected sequentially to form the polygon. Hit the Space/Enter key to finish drawing and close the shape.

In the initial state, the drawn polygon will be on a plane that remains parallel to the ground.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Press the K key to enter Shared-Edge-Polygon Mode. When creating a new polygon, you can connect two points on an existing shape to share edges. Use L to switch the shared edges.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

The result has three views on the right side, allowing independent scaling and translation. You can use the shortcuts Q/A/D/E to fine-tune the polygon position in the Overhead view and adjust its height above the ground using W/S in the Side view.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Polyline

Select Polyline from the toolbar. Click on the canvas to add points, connecting to form a polyline. Press the Space/Enter key to finish drawing.

The result has three views on the right side, allowing independent scaling and translation. You can use the shortcuts Q/A/D/E to fine-tune the polyline position in the Overhead view and adjust its height above the ground using W/S in the Side view.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Keypoint

Click on the edges of an existing polygon or a polyline to add nodes. After selecting these key points, you can annotate (T), delete (Bks), or drag them individually in the three views.

Shortcuts can only adjust the entire polygon or polyline but not the keypoint.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Group

To group multiple results, hold down the Shift key to select them and then press Ctrl / ⌘+G.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

You can remove a specific object from a group by clicking UnGroup in the Results List. To dissolve the entire group, select it and press Ctrl / ⌘+Shift+G.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Measuring Tool

Click on the point cloud data to place the starting point, and click again to add nodes. Press the Space/Enter key to finish measuring.

To delete, select the measuring line and press the Bks key.

Measuring lines are not results; they will not be saved upon exiting the Tool page.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Run Model

It is advisable to utilize a model for quicker object detection.✨ You can click ... below the icon to configure it. Select the model, set the confidence, distance, and model classes as needed, then apply the settings and run.

Click the model icon again. The data canvas and the results list will show the AI annotation results.

You can manually conduct class labeling and true-value assessment on these AI model results. The following sections provide detailed instructions.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

3. Projection & 2D Annotation

By default, annotation results on Point Cloud Data are automatically projected onto 2D Camera Data in real-time.

You can also manually project them by selecting the 3D results and clicking the Projection button. To change the Projection Type and Auto Projection, click ... below the Projection button.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Above the 2D Camera Data list, you can select the result types you want to display.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Due to inaccurate camera parameters, annotators may need to adjust 2D results manually. Double-click on the image to enter 2D Annotation Mode. Then, you can modify the size and position of the results on the canvas.

It's important to note that the modified 2D results will no longer update in real-time with the 3D results. If you want to re-project the new 3D results to 2D, click the Re-projection button.

2D results are independent of each other across different images. Therefore, you may need to adjust the results for each image separately.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

In 2D annotation mode, you can also utilize tools to detect objects directly on the image. Check the Image Tool page for detailed drawing methods.

Clicking the Track Line icon will display the auxiliary line and circle while drawing with the tool. You can adjust the settings for these auxiliary elements by clicking on the ... icon below.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

4. Class & Attributes

After detecting an object, choose its Class and provide the corresponding Attribute information. Press T to show/hide classes and attributes pad.

🚧

Exercise caution when changing an object's class to avoid losing its attributes.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Class

Select an appropriate Class for the object. You can use the search function to find the Class quickly.

Hovering over a class tag with a question mark displays the constraints and standards. A "Class Invalid" error appears if the result fails to meet these limits.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

By clicking the Set as Standard button, you can rewrite the Length, Width, and Height of the selected object to meet the required standards.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Attributes

After selecting the Class, you need to fill in its Attributes, including the General attributes, the LiDAR attributes for the point cloud result, and the Image attributes for the 2D camera result.

🚧

Each 2D result in distinct images has independent Image attributes that should be provided separately.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Attributes for Point Cloud Results

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Attributes for Image Results

As for tracking objects in a Scene (series frames), you can either copy attributes from other objects in the same frame or apply the current attributes to other frames of this object.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Advanced

In the advanced section, you can easily update information for the selected object across all frames, including marking all results as ground truth, switching the object type, merging & splitting, and deleting.

Object Type applies to scene tracking: Fixed only changes position info without altering the box size, while dynamic changes both size and position.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

5. Result Management

Result List

The annotated objects will be automatically numbered and categorized by Class in the Result List. You can click on an object to view all 3D and 2D results or select a specific 3D/2D result within it for editing. Press Esc to cancel selection.

📘

An Example

In the image below, "3D(2825)" represents the 3D result containing 2825 points for Object #69, "C2 Box" represents the Box on the 2nd camera data, and "C3 Rect" represents the Rectangle on the 3rd camera data.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

You can edit the entire Class, including quickly deleting objects within it, toggling visibility, or switching to another Class (the same tool type is required).

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

The results will display different validity statuses. Model-generated results marked "No True-Value," and manually processed results marked "All True-Value."

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

You can also filter results by Class, tool, and frame at the top of the result list.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

📘

Validity, Classifications, and Comments

On the right of the results list, you can switch to the Validity, Classifications, and Comments lists.

These properties and comments are specific to the entire Data or Scene. For more details, you can visit the Annotation Tools Commons page.

Filter Other Object

When the Filter Other Object option in the Toolbar is enabled, other objects will be filtered while selecting or drawing an object.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Result Focus

When the Result Focus option in the Toolbar is enabled, only point cloud data around the selected object (polygon or polyline) will be displayed. You can customize the Focus Range by clicking ....

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

6. Object Tracking in Scene

When annotating a Scene, you can track objects across frames in the Timeline area.

At present, the tracking feature only works with objects that have been annotated using the Cuboid (3D box) tool.

Auto Object Tracking

Start by selecting an object of interest within a labeled frame, usually the default first frame, and configure the object tracking process.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

There are three methods to identify the object and predict its position:

  • Basic Object Tracking Model: Predict the object position in each frame by model.
  • Copy: Duplicate the result to other frames with identical positions.
  • Interpolation Tracking: Calculate the object position in specific frames based on the current results. At least two ground truth values on the timeline are required; otherwise, errors may arise.

You can choose options for applying to the selected object/all objects, tracking in forward/backward frames, and specifying frame number (Increased runtime for more results and frames).

Then, click the Object Tracking button. The tracking results will be applied to Point Cloud and Image Data across frames a few seconds later.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Model & Copy Tracking

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Interpolation Tracking


Review in Batch

After tracking, you can click the Review button to batch review and edit results in all frames.

For Point Cloud Data, you can switch between different views to adjust the results of the selected object, such as moving and scaling the cuboid or flipping the head.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Results of the same object in different frames share the same class and attributes. You can check the Show Attributesoption to modify them quickly.

A "Class Invalid" error will occur on the timeline if the result fails to meet the class limits.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

You can also perform Interpolation Tracking here. Multi-select the frames you want to track (including at least one ground-truth value), or choose all frames, and then click the Interpolation button.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Switch to Camera Data. You can choose different sensors to check results (but modifications should be made in 2D Annotation Mode).

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

🚧

No results for the Camera Data?

You may need to complete the projection operation first.

If there are still no results, it's possible that no objects are within the current camera's field of view or that there are errors in the camera parameter settings.

By clicking the Trajectory button, you can view a trajectory line composed of a series of frames, which aids in analyzing object tracking.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Merge

Select an object (X) and click the Merge icon above the Timeline. Choose the Merge Type and the Target (Y).

  • X Merge From Y: Merge Y to X, keeping the Tracking ID and Class of X and deleting Y.
  • X Merge To Y: Merge X to Y, keeping the Tracking ID and Class of Y and deleting X.

You can either select a target from the dropdown menu or use the icons beside it to click on the object directly in the Point Cloud Data.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

🚧

NOTICE:

The frames of X and Y must not overlap.

X and Y must have the same class.

Objects cannot be merged with themselves.

For example, if result #54 (in frames 1-3) and result #425 (in frames 4-8) pertain to the same object, we can merge them to maintain the consistent Tracking ID and Class. Select result #54 and merge to the target #425. Then, result #54 will be deleted, and its information will be merged into result #425.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Split

Like the reverse of merging, an object can be divided into multiple parts.

Select an object (X) and choose a specific frame. Click the Split icon and set the Class for the new object (Y). X and Y will be separated from the selected frame.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Fold

You can fold frames to check objects' stationarity and annotate across multiple frames rapidly.

🚧

Prerequisite

Folding is only possible when the point cloud's coordinate system doesn't move with the ego vehicle during data collection. Otherwise, errors may occur like this:

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Click on the starting frame and hold down the Shift key to select the ending frame. Then click the Fold button.

All results existing in the selected frames will be combined into the folded frame for display only. You cannot select individual results or edit their class attributes, and the results list will be cleared. (Unfolding will restore the original state.)

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

You can directly create new results on the folded frames and edit class attributes. Upon unfolding, the newly created results will be applied to each frame.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Merging and splitting will be simultaneously applied to each frame within the folded frames. The folded frames can also be further folded with other frames.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Merge

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Split

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

Fold with other frames

🚧

Please noted that folded frames are NOT a distinct result type.

Batch review, switching instance/segmentation mode, submitting, and other task-related operations will automatically split them.

Folding frames may result in excessive points, requiring a sparse display. You can downsample point clouds in the display settings, located in the bottom left corner.

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object

7. Shortcut Keys

annotate data, data labeling, annotation tool, lidar fusion, point cloud, object detection, tracking object