Creating a 3D Gaussian Splat (3DGS) model follows a structured workflow: data capture, dataset preparation, training, and export. A Gaussian Splat model represents a scene using millions of 3D ellipsoids, known as "splats." Each splat contains:
Instead of building mesh geometry, the scene is rendered directly from these splats using GPU rasterisation.
Data capture determines model quality and typically represents the majority of the effort.
Use a high-resolution camera with:
Shoot in 4K (minimum) if recording video. Maintain strong parallax—avoid simple left-to-right pans. Move through the space to create depth variation.
Use UAVs with large sensors and full manual controls. Fly slow, consistent paths and achieve 70–80% image overlap. Capture multiple altitudes for complete geometry.
Manual exposure and focus are essential. Automatic settings frequently cause reconstruction failures.
Images or video frames are organised into a structured dataset ready for processing.
Common software options include:
Tool choice depends on dataset size, GPU capability, scalability requirements, and budget.
Training is the optimisation phase where the system reconstructs the scene.
The process:
The goal is to minimise the difference between rendered views and original images.
During densification:
Training time ranges from minutes to many hours, depending on dataset size and GPU performance. Modern GPUs (NVIDIA, AMD, Apple Metal) dramatically accelerate processing.
Once the data has trained, the 3DGS model can be exported as:
These outputs support measurement, remote inspection, planning, collaboration, VR/AR, and engineering workflows.
For commercial deployment:
Most reconstruction failures stem from capture errors:
Nearly all issues trace back to inconsistent or incomplete data capture.
Successful Gaussian Splat creation depends on:
Master the capture stage first—because no amount of processing can repair fundamentally poor input data.
