LiDAR Annotation Tool
This article describes how to use 2D & 3D Fusion and Frame Series tool. The object tracking is a key-feature of the perception system of autonomous cars and ADASs.
LiDAR Annotation Tool Overview
The main screen of the LiDAR annotation tool consists of the following components
LiDAR Annotation Tool Interface
- Create Box: Create a new cuboid on 3D canvas;
- Translate: On/Off switch Move the cuboid when it is on; Support moving to a certain axis of X/Y/Z;
- Run Model: Run the model to predict results
RGB Camera Data
- To project b-box/2d cuboid on 2d image;
- If it is not a 2D & 3D Fusion dataset, 2d images are used for reference, helping to identify objects that are not clear in 3D.
Point Cloud Data
- Annotate objects according the data annotation guideline;
- AI-assisted tool can precisely make the cuboid fit on the object automatically
- There are many settings that can help annotators better identify objects.
For example, by displaying the intensity value of points, changing the size of points, etc.
- When a cuboid is created on the 3D point cloud data area, minor adjustments are usually made here to make the cuboid fit as closely as possible to the target object.
- All annotation results are displayed here, including Validity, Classification and Class results.
Display data information, including coordinate values and the total number of points.
Click Create (or hit F on the keyboard) on the upper right corner, click 3 time to draw the outline of the object, then a cuboid that fits the object will be generated by the AI-assisted tool.
AI-assisted 3D Cuboid Tool
In most cases, slightly adjust the cuboid in three different views: overhead, side and rear views.
Hit C on the keyboard to change the cuboid facing direction.
Adjust 3D Cuboid in 3 Views Area
It is recommended to use shortcut keys to adjust the position.
Note: Using shortcuts do not reset the scale of three views.
If you're working on a 2D & 3D fusion dataset, Projection and Re-projection button can project a bounding box or cuboid to the RGB images from point cloud data.
You can adjust the bounding box and cuboid in RGB images, click Re-Projection to reset the result.
Project 3D Results to the RGB Image
Xtreme1 platform has preset models for ADAS scenarios -- only by clicking one button to auto-generate results. Yon can also integrate other models into the Xtreme1 platform. See Model Page.
Click Run Model button to apply a model to auto-annotate objects in 3D point cloud data. Click the button again to load results after recognition.
3D Object Detection Model Tool
Running model predicts all targets in point cloud data by default, you can choose preferred classes and confidence value in AI Annotation Setting.
Pre-annotation Model Setting
Click Info button on the lower left corner to show Coordinates, Points of selected scene results.
Click Setting below data info to customize display effects of Image, Point cloud and Result etc.
- 1.Image: Tick checkboxes to show Cuboids, 3D Cuboids, Projected Cuboids and Projected Points
- 2.Point Cloud: Points - Point size in 3D point cloud; Color Height / Intensity - Slide or input value to modify point cloud render color; Click Reset to reset to default value;
- 3.Objects：Switch to show tags on results;
- 4.Utility：Switch to show distance measure and its radius in meter.
Xtreme1 Tutorial Series Episode 2: Annotation of 2D & 3D Fusion Data