For the complete documentation index, see llms.txt. This page is also available as Markdown.

Combined Tools (Cuboid, 3D Polygons, 3D Polylines + 2D Drawables)

Cuboid + Polygon + Brush - combined in a single annotation workflow to segment point clouds

Overview

The Multi-Modal Segmentation Tool is a new point cloud annotation capability that integrates three distinct segmentation methods - cuboid, polygon, and brush - into a single, unified workflow. Annotators can now apply the optimal tool for each geometric region of an object without switching contexts, losing progress, or accepting accuracy trade-offs.

Real-world objects are geometrically diverse. A vehicle has a rectangular body, an irregular roof rack, and ground-contact surfaces that each demand a different annotation approach. Until now, annotators were forced to pick one tool and compromise everywhere else - or spend excessive time applying the wrong method to each part.

What's New

Integrated Tool Switching

Annotators can now transition between cuboid, polygon, and brush modes within a single annotation session on the same point cloud object. No work is lost when switching modes, and all segments contribute to a unified output.

Simultaneous Multi-Tool Operation

All three tools operate on the same point cloud simultaneously. Segments created by different tools are merged into a single, coherent annotated object, preserving geometric accuracy across every region.

Contextual Tool Optimization

Each tool is designed for the annotation challenge it handles best. Cuboids provide fast, precise coverage of rectangular geometry. Polygons handle irregular outlines with accurate trace-based boundaries. Brush delivers pixel-level refinement for terrain contact and complex surface edges.

Tool Reference

Tool

Best For

Use in Multi-Modal

Cuboid

Rectangular bodies, geometric shapes

Vehicle chassis, building facades

Polygon

Irregular outlines, complex contours

Roof racks, cargo attachments

Brush Sphere

Fine-detail refinement, terrain edges

Ground contact points, irregular surfaces

Example Workflow

The following illustrates how the three tools work together when annotating a construction vehicle:

Step 1 - Brush: + icon (class level): creates a new annotation instance. Tooltip: "Add annotation". Always visible on the class row.

  • Pencil icon (annotation level): edits an existing annotation. Tooltip: "Edit with".

    • Clicking the pencil icon beside an annotation opens a palette labelled "Edit with" listing all drawables configured for that class.

    • Selecting a drawable enters add/remove-point editing mode for the existing annotation using the chosen tool.

    • If exactly one drawable is configured, the pencil icon activates it immediately without showing a palette.

    • The edit palette is sourced from the same project-setup configuration as the create palette. No hardcoded tool lists exist in either.

Step 2 - Polygon: + icon (class level): creates a new annotation instance. Tooltip: "Add annotation". Always visible on the class row.

  • Pencil icon (annotation level): edits an existing annotation. Tooltip: "Edit with".

    • Clicking the pencil icon beside an annotation opens a palette labelled "Edit with" listing all drawables configured for that class.

    • Selecting a drawable enters add/remove-point editing mode for the existing annotation using the chosen tool.

    • If exactly one drawable is configured, the pencil icon activates it immediately without showing a palette.

    • The edit palette is sourced from the same project-setup configuration as the create palette. No hardcoded tool lists exist in either.

Step 3 - Cuboid: Draw a cuboid around the vehicle's main rectangular body. Fast, precise, and ideal for geometric shapes.

  • Leverage all cuboid features to get fast and accurate segmentation. Features that help reduce the time per object and enhance accuracy: interpolation, film reel, autogrounding, static dimension etc.

The result is a fully segmented point cloud built with the optimal tool for each geometric challenge - completed in one continuous workflow, with no context switching.


Configured Drawables

Smarter Annotation Palette

  • The drawable palette - the menu annotators use to choose their labeling tool - shows only the tools configured for the class being annotated.

  • When a class has only one tool configured, the palette is skipped entirely. The tool activates immediately and a brief confirmation strip tells the annotator which tool is active.

Edit With Any Configured Tool

  • Annotators can edit an existing annotation using any drawable configured for that class - not just the tool that originally created it. This means an annotation created with Polygon can be refined using Brush sphere, and vice versa, without needing to re-annotate from scratch.

  • The edit and create palettes are sourced from the same project configuration, ensuring consistency.

Visually distinct create and edit actions

  • The class list panel uses distinct icons to differentiate creating a new annotation from editing an existing one. The + icon creates. The pencil icon edits.


Availability & Compatibility

Availability: This feature is generally available in the current platform release.

Supported formats: All standard point cloud formats currently supported by the annotation platform (pcd and las).

Existing projects: Fully compatible with existing annotation projects. No migration required.

Access: Available to all annotator roles. No additional permissions or feature flags required.


Why It Matters

As autonomous systems encounter increasingly diverse real-world objects - construction equipment, unusual vehicle configurations, complex urban structures - training data must capture geometric complexity accurately. Single-tool annotation workflows cannot keep pace with this diversity.

Multi-Modal Segmentation addresses this directly:

  • Accuracy improves because each region is annotated with its geometrically appropriate method.

  • Speed increases because annotators are no longer spending excessive time forcing the wrong tool onto a task.

  • Workflow continuity is preserved - no saving, switching, or re-importing required between tool modes.

  • Training data quality scales with object complexity, rather than being bottlenecked by it.

Last updated