Develop · Build
How it came together.
- 01
Leading with the failure heatmap
The draft opens with the PSNR-by-engine × scan heatmap, failure cells highlighted in red, before the per-engine averages. Robustness comes before the leaderboard.
Research - 02
Five difficulty tiers as the spine of the eval
The eval runs on five tiers, easy through extreme, from a clean studio booth to a commercial office, each chosen for a distinct failure mode: texture sparsity, specularity, scale ambiguity, lighting variance.
Research - 03
Reproducibility as a second contribution
The paper ships with the eval harness, the scan-difficulty rubric, and the per-engine launch scripts, so someone else can run the same framework against their own engine release.
Data Pipelines - 04
A real convergent result on a reflective scan
The companion paper on reflective interiors reached its first convergent training run on a real, highly reflective showroom scan, with a measurable gain from the reflection branch alone. Residual ghost geometry in mirrors is reported as a limit of single-pass reflective reconstruction, not smoothed over.
Research