Skip to main content
The Gaussian Splat viewer renders 3D scene reconstructions using WebGPU-accelerated Gaussian splatting. Navigate photorealistic 3D environments and place annotations directly in the scene — no separate 3D modeling tools required.

​
Overview

Gaussian splatting is a 3D scene reconstruction technique that represents environments as collections of 3D Gaussians. Each Gaussian stores position, covariance, color, and opacity, and the renderer composites them into photorealistic views from any camera angle. Unlike mesh-based representations, Gaussian splats capture fine details like foliage, reflections, and semi-transparent surfaces. In Avala, you can upload Gaussian Splat files to a dataset and annotate them directly in the viewer — placing 3D cuboids, assigning classification labels, and organizing annotations in a scene hierarchy.

​
Viewer Layout

The viewer is organized into five regions: Explore the reconstructed scene using standard 3D camera controls:
When annotating, the viewer switches between navigation mode and annotation mode. In navigation mode, mouse inputs control the camera. In annotation mode, mouse inputs interact with annotations. Toggle between modes using the control at the top center of the viewer.

​
WebGPU Rendering

The viewer uses a WebGPU-accelerated rendering pipeline optimized for real-time Gaussian splatting.

​
Pipeline Architecture

​
Browser Support

The Gaussian Splat viewer requires WebGPU. Unlike the point cloud viewer, there is no WebGL fallback for Gaussian splatting — the GPU compute pipeline is essential for real-time splat sorting and compositing. Use Chrome or Edge 113+ for the best experience. WebGPU features are under active development and may not be available in all environments.

​
Annotation

Place annotations directly in the 3D Gaussian Splat scene using the toolbar at the bottom of the viewer.

​
3D Cuboids

Draw three-dimensional bounding boxes around objects in the scene:
  1. Select the Cuboid tool from the toolbar
  2. Click in the scene to place the cuboid center
  3. Adjust position, dimensions, and rotation using the gizmo handles
  4. Assign a label from your project’s label taxonomy
Cuboid placement is depth-aware — the viewer uses the Gaussian Splat depth buffer to position annotations at the correct 3D location in the scene.

​
Classification Labels

Apply scene-level or object-level classification labels:
  • Scene classification — Label the entire scene (e.g., indoor, outdoor, parking lot)
  • Object classification — Label individual cuboid annotations with object classes and attributes

​
Scene Hierarchy

The left panel displays a tree view of all annotations, organized by label class. Use the hierarchy to:
  • Select annotations by clicking their name
  • Toggle annotation visibility
  • Reorder and group annotations
  • Inspect annotation properties without navigating to them in the 3D view

​
Edit History

Full undo/redo support is available via the top-center controls or keyboard shortcuts: Every annotation action — creation, deletion, move, resize, label change — is recorded in the edit history.

​
Use Cases

3D Scene Understanding

Annotate objects in reconstructed indoor and outdoor environments for 3D object detection and scene graph generation.

Novel View Synthesis

Label objects in Gaussian Splat scenes to create training data for view synthesis models that need semantic awareness.

Spatial AI Training Data

Build labeled datasets for embodied AI agents that need to understand and navigate 3D environments.

Digital Twin Annotation

Annotate reconstructed real-world environments for digital twin applications, facility management, and simulation.

​
Next Steps

3D Cuboid Tool

Detailed guide for placing and adjusting 3D bounding boxes.

3D Point Cloud Viewer

GPU-accelerated point cloud visualization with six rendering modes.

Data Types

Overview of all supported data modalities and their capabilities.

Multi-Sensor Viewer

Synchronized playback of camera, LiDAR, radar, and IMU data.