Show HN: High-fidelity, compact, and real time rendering of university campus (hoanh.space)

A creator has released an interactive, high-fidelity 3D model of Aalto University that runs in real time in a web engine, demonstrating a practical pipeline built on recent advances in Gaussian Splatting. The scene was captured with a DJI Mini Pro 5 drone, photogrammetrically aligned with COLMAP, trained into a Gaussian Splat representation using gsplat with MCMC, compressed with SOGS, and rendered via the PlayCanvas engine. The result is a compact, photorealistic digital twin that can be explored in a browser, showcasing how neural rendering techniques can be deployed outside research prototypes. For the AI/ML community the project is a clear proof-of-concept that Gaussian Splatting plus targeted compression makes large, real-world scenes interactive and web-deployable. Key technical takeaways are the hybrid pipeline—classical structure-from-motion (COLMAP) feeding a learned Gaussian representation, MCMC-based training for stability or refinement, and SOGS compression to meet real-time constraints—integrated into a lightweight runtime (PlayCanvas). This demonstrates practical paths for scalable neural rendering, streaming digital twins, AR/VR content, and further work on compression, level-of-detail, and browser GPU optimizations.