SplatPrep Studio: 360° Video/Panorama to Multi-View Training Data WIP
SplatPrep Studio addresses a persistent challenge in 3D reconstruction workflows: efficiently generating comprehensive multi-view datasets. The application transforms 360°/180º equirectangular mono or stereo video or high-resolution equirectangular panoramas into perspective camera views suitable for both 3D Gaussian Splatting (3DGS) and traditional photogrammetry pipelines. This is expanding upon work I did in 2010 (https://www.petermorse.com.au/2010/01/09/fulldome-virtual-cameras-in-mawsons-huts/)
Of course, people thought about this stuff in the C19th – it’s just about how you do it. Look at the seminal work of Paul Bourke (http://www.paulbourke.org/panorama/) – and the top-notch work of Andrew Hazelden (https://github.com/AndrewHazelden) and Olli Huttunen (https://www.youtube.com/@OlliHuttunen78/videos).
Traditional dataset capture requires methodical photography with careful attention to overlap percentages, baseline distances, and lighting consistency. This process becomes particularly tedious for complex scenes or when operator shadows must be avoided. 360° video capture simplifies acquisition—a single pass captures all viewpoints simultaneously—but introduces the challenge of extracting useful perspective views from spherical projections.
SplatPrep Studio provides a complete pipeline for this transformation. The workflow begins with loading equirectangular footage and planning camera positions through an interactive preview overlay. The application then renders perspective views via Blender integration (supporting both headless batch processing and an interactive addon workflow), filters results using automated quality metrics, and exports the final dataset with camera pose information compatible with COLMAP and similar reconstruction tools.
The camera placement system supports multiple tilt groups with independent configuration. Users can define level, downward-angled, and upward-angled camera arrays, specifying camera counts per tilt level. The system includes automatic operator occlusion avoidance, identifying regions where the camera operator appears in the footage and excluding camera placements that would capture these areas. Notably, the application supports different frame cadences and offsets for Gaussian Splatting versus photogrammetry—the former benefits from dense overlap whilst the latter requires wider baselines for robust geometry reconstruction.
Implementation centres on a cross-platform desktop application. The architecture separates concerns across distinct tabs: Project configuration and camera planning, rendering orchestration, quality assessment, and COLMAP integration. The application persists project state, enabling iterative refinement of camera placements and rendering parameters.
Blender integration operates through two mechanisms. For batch processing, the application generates Python scripts executed via Blender’s headless mode, positioning virtual cameras within a 360° environment sphere and rendering perspective views at specified resolutions.
Quality filtering employs computer vision techniques to identify optimal frames. Blur detection helps exclude poorly focused images, whilst exposure analysis identifies over- or under-exposed regions. This automatic filtering proves essential when working with hundreds of rendered frames, reducing manual review overhead considerably.
Finally – the app works with a self-contained version of COLMAP. This simplifies image registration and training for both Gaussian Splatting and Photogrammetric workflows downstream.
The project will be released as shareware. Details TBA.
Processing performance is being calibrated. It’s not too bad – depends on your hardware. Can be optimised.
SplatPrep Studio demonstrates that 360° video and panoramas shot on consumer hardware, often dismissed as a novelty format, provides genuine utility for professional 3D reconstruction workflows when paired with appropriate processing tools. The application will be available on GitHub with full source code and documentation- and binaries on Gumroad in 2026.
