COLMAP demo datasets as TypeCV benchmarks

The COLMAP demo datasets (south-building, gerrard-hall, person-hall, graham-hall) ship three things: the images, the COLMAP database with SIFT keypoints, descriptors, raw matches and verified inlier matches, and a bundled sparse model from COLMAP's full pipeline. scripts/colmap-datasets/colmap_dataset.py turns one dataset into a controlled benchmark:

1. export writes the TypeCV sequence JSON from the database: the exact COLMAP SIFT keypoints and raw matches, with the bundled model's refined intrinsics held fixed. COLMAP's +0.5 pixel-center convention is shifted to TypeCV's integer centers for keypoints and principal point. A subset (--images, --step, --window) keeps a run inside the current mapper's limits.
2. reference summarizes the bundled model on the same image subset, and with --run-mapper runs pycolmap's incremental mapper on a copy of the same database with the same fixed intrinsics. That run isolates the mapper: both systems see the same features and the same putative matches.
3. scripts/reconstruct-sequence.mjs (CPU reference path) or scripts/reconstruct-sequence-gpu.mjs (TypeShade path in Chromium) reconstructs the sequence.
4. compare similarity-aligns the TypeCV cameras to each reference on their common images and reports rotation and center agreement, registration, point and observation counts and elapsed time.

The references are not ground truth. Both systems share the same features and matches, so a systematic matching error appears in both. Counts and reprojection errors use different accepted observation sets and are not accuracy claims. The bundled model used COLMAP's own intrinsics refinement over all 128 images; TypeCV and the mapper run hold those refined intrinsics fixed, which favors both.

Reproduce

Use Python 3.12+ and Node 22.18+. pycolmap 4.2.1 is pinned in scripts/requirements-sequence.txt.

npm ci && npm run build
python -m venv .venv-sequence
.venv-sequence/bin/python -m pip install -r scripts/requirements-sequence.txt
D=.venv-sequence/bin/python scripts/colmap-datasets/colmap_dataset.py
$D fetch south-building data/colmap
$D export data/colmap/south-building data/sb16/sequence.json --images 16
echo '{"maxErrorPx": 4, "ransacIterations": 750, "seed": 42}' > data/sb16/options.json
$D reference data/colmap/south-building data/sb16/reference.json \
  --sequence data/sb16/sequence.json --run-mapper data/sb16/colmap --fixed-camera
node scripts/reconstruct-sequence.mjs data/sb16/sequence.json data/sb16/typecv data/sb16/options.json
$D compare data/sb16/typecv data/sb16/sequence.json data/sb16/reference.json data/sb16/comparison.json
# TypeShade path (software WebGPU by default; TYPECV_HARDWARE_GPU=1 for a physical adapter):
node scripts/reconstruct-sequence-gpu.mjs data/sb16/sequence.json data/sb16/typecv-gpu data/sb16/options.json
$D compare data/sb16/typecv-gpu data/sb16/sequence.json data/sb16/reference.json data/sb16/comparison-gpu.json

maxErrorPx 4 matches COLMAP's default filter at the source resolution (3072 × 2304). Use --max-edge on export to scale coordinates and intrinsics together when a run must stay at a lower analysis resolution.

Recorded results: south-building (2026-10-19, this container, no physical GPU)

All runs use the same COLMAP SIFT keypoints and raw matches and the bundled refined intrinsics held fixed. maxErrorPx 4, 750 RANSAC trials, seed 42. The pycolmap mapper is the CPU wheel. TypeCV CPU runs are the synchronous reference path on one core; the GPU run is Chromium's SwiftShader software WebGPU, so its time measures correctness and structure, not acceleration.

Records: south-building-16.json, south-building-32.json, south-building-6-gpu-software.json, south-building-16-persistent.json, south-building-32-persistent.json.

The persistent-landmark rows (the mapper keeps its landmarks across registrations and triangulates only completed tracks, with a closed-form 3×3 inverse in CPU BA) were timed while SwiftShader tests ran on the same container, so they are upper bounds. Their profile puts the remaining CPU time in bundle adjustment (local passes still carry every landmark) and in eight-point pair verification.

Observed: camera agreement with COLMAP's mapper is within 0.06° on both subsets. TypeCV's CPU mapper is 18× slower at 16 images and 33× slower at 32 images, and its time grows about 4.9× for 2× images. The landmark counts differ because the two mappers accept different tracks (TypeCV's triangulation angle threshold is 0.5°, COLMAP's default is 1.5°); they are not a quality ranking. The profile of the 16-image run puts the time in repeated track triangulation (nullspace), bundle adjustment and garbage collection: the mapper re-triangulates every track after each registration. That is the first mapper change to make.