What is 3D Gaussian splatting?

A way of turning ordinary photographs of a real place into a 3D scene you can move through in a web browser. Rather than modelling the room as surfaces, it stores millions of tiny translucent blobs of colour, each stretched and tinted to match what the cameras saw. Drawn together they reconstruct the view from angles nobody actually photographed, fast enough to walk through in real time on ordinary hardware.

Can 3D Gaussian Splatting run inside a mobile web budget?

Yes. The applied benchmark closes the gap between the Kerbl et al. 2023 method paper and what actually deploys to a Safari-iOS budget; Barnes Vantage is the shipped web-native viewer, live in the BARNES Dubai property-search flow.

Which 3D Gaussian Splatting engine actually deploys to production?

Across a 26-engine benchmark, PSNR/SSIM/LPIPS evaluation is in progress, with an early working hypothesis that the quality winner will not be the speed winner. The engine that ships is the one that survives a production Safari/iOS budget.

Why can't 3D Gaussian Splatting reconstruct mirrors and glass?

Splatting reconstructs surfaces from photographs, but a mirror or floor-to-ceiling glass wall is a reflection, not a surface. Pipelines ghost the reflected room into the wrong volume or hollow the geometry behind the glass. Mirrors, glass and chrome are the dominant artefact class on hard scans. The primary bet is capture-side handling at the photogrammetry layer, though a render-side reflection branch is also in progress and has reached a first convergent result on a real reflective showroom scan.

Does a domain-tuned LLM beat a base model with RAG?

For a narrow vertical (Dubai luxury real estate), yes, on terminology and named-entity recall. Barnes Dubai LLM v8 is a QLoRA fine-tune trained on 15,369 instruction pairs sourced from three national property registries (Dubai DLD, UK Land Registry, France DVF), Singapore URA transactions, and World Bank economic indicators.

Back to portfolio

Research

From paper, through benchmark, to shipped.

Seven lines of work, each starting from published research, passing through a benchmark or implementation of my own, and ending at something running in production. Open a thread for the papers it rests on and what it found, including the findings that did not go the way I expected.