This Gaussian Splatting scene captures the greenhouses of the Botanical Garden in Kiel ( https://www.uni-kiel.de/en/botanic-gardens-kiel ). The capture process utilized a DJI Osmo Action 5 Pro recording in 4K 4:3 video, with approximately five to seven minutes of footage per room. Frames were extracted using Sharp Image Extractor ( https://sharp-frames.reflct.app/ ) and subsequently posed using RealityScan, resulting in eight individual rooms with fully posed cameras. Each room was exported as an alignment component and merged into a master scene comprising nearly 9,000 poses.
Training was conducted in LichtFeld Studio ( https://lichtfeld.io ). The pipeline began with training all 9,000 poses into a 20M SH1 splat, followed by individual training runs for each room, and targeted passes for every connecting doorway with cameras positioned on both sides. These individual outputs were then combined using splat-merge ( https://github.com/simonbethke/splat-merge ).
For optimization and performance, a 1M SH1 scene was trained to serve as the lowest level of detail (LOD). Collision data was generated by cleaning the 20M splat using SuperSplat ( https://github.com/playcanvas/supersplat ) and deriving the collision geometry from that processed mesh. The final package was processed with splat-transform, uploaded, and published.