03·3D & Spatial
3DGS Research Pipeline
20 Gaussian-splat engines integrated into one benchmark harness, 6 evaluations run so far. Which one actually survives a real luxury-interior scan, not just the leaderboard.
ResultA reproducible benchmark harness: around twenty engines wired, 6 evaluations run so far, a scan-difficulty taxonomy and per-engine failure modes, that the academic write-up draws on.
Discover
Why it exists.
3D Gaussian Splatting moved fast between 2023 and 2026. New engines shipped every quarter, each claiming higher PSNR on their preferred benchmark scenes. The applied question, which one actually survives a luxury real-estate brief, on a real scanner, in a real building, sat unanswered.
This pipeline answers it. Around twenty Gaussian-splat engines wired into one harness, the first evaluations run across the difficulty range from a clean scan-studio booth through a mirror-walled marble lobby to a city-scale neighbourhood capture, the full matrix in progress. Not just average quality, robustness, with failure modes per engine logged, because an engine that crashes on one scan in five is unusable in production regardless of its PSNR on the other four.
The pipeline also covers the unbuilt case: novel-view synthesis via video diffusion, monocular depth lift, and the hybrid stack that bolts those onto a real scan when only half a building has been constructed.
StackCUDA · gsplat · PyTorch · COLMAP · Video diffusion · Monocular depth
Define
The brief.
Answer which Gaussian-splat engine actually survives a luxury real-estate brief on a real scanner, not which one wins average PSNR on clean studio scenes.
Develop
What it took
Skills behind it.
Primary discipline plus the support stack.
Skills demonstrated
- ResearchPrimary
Around twenty 3DGS engines wired into one harness spanning studio scans to city-scale capture, with the first evaluations run on PSNR, SSIM and LPIPS and the full matrix in progress.
- 3D & SpatialSecondary
Novel-view synthesis and monocular depth paths for unbuilt properties stitched into the same pipeline.
- AI & Machine LearningSecondary
Video-diffusion novel-view models evaluated against COLMAP-grounded baselines.
- Data PipelinesSupporting
Reproducible eval harness across PSNR, SSIM, and per-engine failure modes.
Develop · Build
How it came together.
- 01
Three difficulty bands, not one leaderboard
Most 3DGS benchmarks score against a small set of scene types, usually clean studio scans where every modern engine clears thirty dB. Split the eval into seven scan-difficulty bands from Easy Scan-Studio through Extreme Commercial-Office, so engines that excel at clean interiors but fail on mirrored marble are scored honestly. Choice of engine matters more than expected at the easy end too.
Research - 02
Failure mode is a first-class metric
Engines do not just produce worse images on hard scenes, they crash, produce floaters, or output garbage point clouds that crash downstream rendering. Logged every fail as a red cell in the per-engine × per-scan heatmap. The write-up leads with this matrix, not the PSNR scores, because robustness is what a production team buys.
Data Pipelines - 03
Novel-view synthesis as a parallel path
Off-plan luxury properties have no scan to start from, the building has not been built. Ran video diffusion and monocular depth lift in parallel to the scanned pipeline, so the same viewer can show a real apartment and an AI-imagined one without two separate codebases. The eval covers both paths against COLMAP-grounded baselines.
AI & Machine Learning - 04
Cleaned vs raw as a routine comparison
Floaters and stray gaussians make scans look worse than the underlying capture warrants. Every scan in the eval was run twice, raw output and after a cleaning pass, to quantify how much of the quality delta is the engine versus the post-processing. Cleaning recovers roughly eight per cent of noisy gaussians on average, and changes which engine wins on the hardest scenes.
3D & Spatial
Develop · Forks
Decisions on the record.
The few calls worth defending. Each one is a fork; the other branch would have been a different project.
Decision · 01
Why gsplat over the reference gaussian-splats-3d implementation
gsplat is the CUDA-native reference path the field is now standardising on, faster training, cleaner gradient flow, and a maintained Nerfstudio integration. The original gaussian-splats-3d implementation was the right starting point in 2024 but it lagged on rasteriser fixes by the time the benchmark started.
Decision · 02
Why RRF for fusion, not a learned ranker
Reciprocal Rank Fusion has no training data dependency and no per-engine weights to tune. With twenty engines across the benchmark the cost of a learned ranker (label collection, train/val splits, drift) didn't earn its keep over a parameter-free fuser that already separates the top tier.
Decision · 03
Why write it up as a paper, not just ship it
The benchmark is applied-3DGS in luxury real-estate visualisation, the gap was not a method, it was an evaluation harness paired with deployment context. Writing it up as an academic paper forces the rigour and makes the work citable, next to the engines it benchmarks.
Operate the result
Drag to compare.
Same scan, two states. Move the divider with the mouse, touch, or arrow keys.
RawCleaned
Deliver
What shipped.
A reproducible benchmark harness: around twenty engines wired, 6 evaluations run so far, a scan-difficulty taxonomy and per-engine failure modes, that the academic write-up draws on. The full engines-by-scans matrix and the papers are in progress. A relighting track now runs alongside the benchmark: an interactive viewer with movable light fixtures and real soft cast shadows, plus a delighting and relightable-retrain line still in progress.
Being written up as an academic paper, in progress. The eval framework continues to run on new engines as they ship, adding a row to the heatmap is a one-day operation now. The pipeline directly informs which engine ships behind the Splat Viewer for each property type.
By the numbers
Engines integrated
Evaluations run
benchmark in progress
Research origin
The paper-to-product thread this project sits inside. The full chain (paper, benchmark, ship) lives on the research index.
- 3D Gaussian SplattingSee thread
Real-time radiance fields applied to luxury real-estate visualisation.
- Mirrors and glass captureSee thread
The open problem Gaussian Splatting still hasn't solved for high-end interiors.
Frames
Selected from the work.
21 frames
01Cleaning Impact 02Cleaning ROI 03Decision Matrix 04Efficiency Comparison 05Quality Ranking 06Raw Vs Cleaned 07Speed Vs Quality 08Benchmark Dashboard 09Effect Size Matrix 10Engine Boxplots 11Engine Ranking 12Radar Analysis 13SSIM LPIPS Scatter 14Win Loss Matrix 15Cleaning Showcase 16Comparison 17Engine Comparison 18Gallery 19Leaderboard 20PSNR Heatmap 21Summary