3D Point Cloud and Episodes
This article about the new labeling interface for 3D Point Clouds in Supervisely that introduces a significantly enhanced workflow, offering extended functionality and improved usability.
The 3D Point Cloud labeling tool in Supervisely is designed for visualizing, annotating, and managing complex 3D data collected from sensors such as LiDAR and RADAR. It supports key tasks like object detection and segmentation across static scenes and sequential episodes, making it ideal for applications like autonomous driving.
The latest version introduces a completely redesigned interface that unifies both single-frame and episode-based workflows. It brings a more streamlined and powerful experience with features such as:
AI-assisted tools for faster and more accurate labeling
Interactive 3D Object Detection
3D Point Cloud Ground Segmentation
3D Cuboid Tracking
Auto Labeling
Synchronized 2Dā3D annotation using photo context images
Timeline navigation for working with sequential frames
Flexible, resizable UI layout tailored to your workflow
Definitions Panel for convenient class management and quick object editing
Advanced settings for customizing visual styles and display preferences
Together, these enhancements provide an integrated and efficient workspace for working with large-scale 3D datasets.
Difference between 3D Point Cloud and 3D Point Cloud Episodes:
3D Point Cloud: A static representation of a scene captured at a single moment in time.
3D Point Cloud Episodes: A dynamic representation consisting of multiple point clouds collected over time, enabling the analysis of movement and change in the scene.
1. 3D AI Assistant
Supervisely's 3D AI assistant is a universal tool for automating 3D point cloud labeling. It covers all types of labeling scenarios for 3D point clouds: 3D object detection, ground segmentation, 3D cuboid tracking, transfer of 2D annotations from photo context images to original 3D point clouds. This tool is class-agnostic - it means that it works with any type of objects regardless of their shape and point density.
Interactive 3D Object Detection
Select smart tool in left side bar and circle target object. It will automatically generate a 3D cuboid around the selected object.
3D Point Cloud Ground Segmentation
Detects and annotates the ground level in the 3D scene.
Fits a horizontal surface through point clusters and creates a flat figure with a
groundclass.Useful for scene normalization and filtering.
Click on auto labeling tab and press "Ground segmentation".
3D Cuboid Tracking
After creating an annotation in one frame, the assistant can automatically propagate it across subsequent frames.
Helps label dynamic objects in sequential datasets with minimal manual input.
Uses a dedicated tracking panel, reusing logic from the video tool.
Auto Labeling
Automatically detects and annotates objects using pre-trained models.
Simplifies the process of placing cuboids or segmenting regions of interest in the scene.
Click on auto labeling tab and enable "Highlight object by click" option, then select manual cuboid tabeling tools in left sidebar and set cuboid on target object.
2D to 3D Projection
The photo context panel is now an interactive part of the 3D labeling workspace.
You can annotate context images directly using standard image labeling tools. These annotations are automatically synchronized with the 3D space and become part of the same object instance. 2D and 3D annotations now coexist at the same level ā edits or creation in one view are instantly reflected in the other. This improves labeling precision and scene understanding, especially when certain features are more visible in 2D.
The system seamlessly combines 2D and 3D perspectives in a single environment ā no need to switch tools or views.
Additional capabilities:
2D masks created on photo context images can be automatically converted into 3D geometry.
Converted figures are visualized directly in the point cloud view.
Currently, only masks are supported. Support for 2D bounding boxes is coming soon.
Click on a photo context image, draw a 2D mask, go to the Auto Labeling tab, and press "Create 3D objects from 2D object on camera."
2. Timeline Support
A full timeline component has been added, similar to the one used in video annotation tools:
Enables navigation across sequential 3D point cloud frames (episodes).
Supports annotation and review of dynamic scenes (episodes) across frame sequences.
Provides a comprehensive overview of frame availability, object presence, and annotation density.
3. Modular and Resizable UI Layout
The new interface allows full layout customization:
Panels such as photo context, camera views, and definitions can be moved and docked anywhere.
Users can arrange the workspace to fit their own workflow and screen space.
This flexibility improves usability and efficiency during annotation.
In addition to repositioning view panels, the Settings panel provides advanced customization options ā such as adjusting cuboid thickness, customizing class appearance, controlling point cloud display settings, toggling object IDs, and more.
4. Definitions Panel
The Definitions panel is now available in the 3D interface, as in image and video tools:
Provides quick access to classes, tags, tool settings, and object styles.
Helps manage large taxonomies and maintain consistency across projects.
Editing
To change the class of a selected object:
Click Select Figure tool.
Select the object in any of the view panels.
In the Definition panel, in the row of the selected class, click the mini-icon with two arrows to change the class.
Summary
The updated interface for 3D Point Cloud annotation combines powerful capabilities:
Integrated 2D and 3D annotation tools
Time-based navigation and frame control
Modular UI layout with dockable panels
Built-in AI Assistant for autolabeling, tracking, and segmentation
It offers a complete workspace for multi-modal annotation with high accuracy and scalability. Whether working with static point clouds or dynamic 3D sequences, the new tool provides clarity, control, and performance required for modern annotation workflows.
Last updated
Was this helpful?
