3dtiled-to-3dtilesBenchmarks
Reproducible conversion benchmark over the repo's own converters, and how it relates to the fastest-tiler ambition.
Why this exists
The TODO / Future work fastest-tiler research note flagged that no
neutral benchmark existed — not against PotreeConverter 2.0 / untwine, and not even run-over-run
against ourselves. This page is the first step: a pinned, reproducible harness measuring our own
converters on fixed public datasets, so any performance change (e.g. wiring the WorkerPool
multicore tier, or the WebP-instead-of-pngjs decode swap) has a before/after number.
Running it
node scripts/benchmark-converters.mjs # all datasets
node scripts/benchmark-converters.mjs --only potree,radSources are mirrored once into .bench-cache/ (download time reported separately and cached across
runs), then the offline convert() is timed cold per run. Results print as a markdown table and
are appended as dated JSON under benchmark-results/.
First results (2026-08-22)
Machine: 12th Gen Intel Core i7-12800H × 20 threads · Node v22.22.3 · Windows 11.
| Dataset | Format | Source MB | Convert s | Output files | Output MB | Throughput |
|---|---|---|---|---|---|---|
| potree1-lion-takanawa | potree | 4.9 | 0.75 | 296 | 7.8 | 0.36 M points/s |
| potree2-lion | potree | 44.2 | 3.92 | 1741 | 111.9 | 1.01 M points/s |
| sog-skatepark | sog | 174.4 | 21.14 | 2055 | 165.8 | 8.3 MB/s |
| rad-tastier-500k | rad | 12.2 | 0.53 | 12 | 9.9 | 0.94 M splats/s |
Reading the numbers:
- ~1 M points/s (Potree 2.0) on a single core (the
WorkerPooltier isn't wired — see Performance) vs PotreeConverter 2.0's published ~9 M points/s for a different job: it builds an octree from raw unordered points; our converter transcodes an existing octree. The honest comparison for the fastest-tiler ambition is our future raw-input tiler vs PotreeConverter/untwine — this table is the regression baseline for the transcode path. - SOG is decode-bound (WebP → DataTable via WASM, then SPZ re-encode): 8.3 MB/s of source. This is the path with the most multicore headroom.
- RAD is nearly free (0.53 s for 500 k splats) — column decode + SPZ encode, no image codecs.
External tilers (context, not apples-to-apples)
| Tool | Job | Their reported speed |
|---|---|---|
| PotreeConverter 2.0 | raw points → Potree octree | ~9 M pts/s uniform / ~6 M pts/s blue-noise (published paper) |
| untwine | raw points → COPC | 2–5× faster than pdal translate past 500 M pts (own claim) |
| MIERUNE point-tiler | LAS/LAZ/CSV → 3D Tiles 1.1 | Rust/Rayon; no published numbers — worth measuring |
None of those transcode already-tiled formats (this repo's job today), and none of their numbers are third-party-verified — see the TODO doc's research note. When the raw-input tiler direction is scoped, these become the targets and this harness grows a raw-input track (same datasets as their READMEs use, e.g. Autzen / CA13).