Configuration

One YAML file per experiment, in projects/<name>/configs/. Where it sits is what says which project it belongs to. It is the whole description of a run: the same file, the same seed and the same photographs give the same result. Unknown keys are refused rather than ignored, so a misspelling fails at the first stage instead of silently doing nothing.

Paths inside calibrate.images, calibrate.intrinsics and colmap.precomputed are resolved against the dataset directory, so a config works wherever the data root is mounted.

name: cpu                  # the run directory; defaults to the file's name
seed: 0                    # every RANSAC and every sampler

sfmkit new writes one of these, commented, as a starting point. A section left out is a stage sfmkit run walks past: no localize means no query to place, no colmap means nothing to be scored against.

calibrate

Where the intrinsics come from. Exactly one of the three.

Key	Default
images	—	glob of chessboard photographs
pattern	[9, 6]	inner corners of that chessboard
intrinsics	—	path to a saved 3×3 K
exif	false	K from the photographs' 35 mm equivalent focal length
sensor_aspect	[4, 3]	the whole sensor's aspect, when the photographs are a crop of it

calibrate:
  exif: true               # 26 mm equivalent on a 4:3 sensor: f = 3029 px
  sensor_aspect: [4, 3]    # the phone's whole sensor; the photos are a 16:9 cut

Digital zoom

A phone reports the equivalent focal length of the lens, not of the crop it saved. A photograph taken with digital zoom therefore reads short — 17% out on the one photograph of this set that had it, which is why that photograph is no longer in it. DigitalZoomRatio is read and applied when the file records it.

sfm

Matching, verification and reconstruction.

Key	Default
images	[]	the photographs to reconstruct from, without extension
reference	—	the camera the model is anchored to, and that comparisons align on
max_keypoints	2048	SuperPoint keypoints per photograph
exhaustive	true	every pair; false pairs each photograph with the reference only
device	auto	auto, cpu or cuda for matching
ransac_threshold	4.0	px, for the fundamental matrix
ransac_iterations	1000
min_inliers	20	below this a pair is dropped
min_track_length	2	observations a track needs
min_triangulation_angle_deg	1.0	below this a point is too ill-conditioned to keep
max_reprojection_error	8.0	px, for filtering observations
pnp_threshold	8.0	px, registering a new camera
min_pnp_correspondences	12	below this a camera is not registered
bundle_solver	schur	schur or scipy: the same problem, one about 58× faster

localize

The query photograph, kept out of the reconstruction.

Key	Default
query	—	the photograph to place, e.g. Img_Old
refine	camera	after RANSAC: none, pose, or camera (K too, when K was estimated here)

colmap

The model to score against: copied, or computed.

Key	Default
precomputed	—	a saved model directory to copy; excludes matches
matches	—	colmap (its own features) or sfmkit (this run's verify/)
camera	self	self, COLMAP calibrates; fixed, it takes calibrate's K

The query always gets a camera of its own, calibrated by COLMAP with its principal point free — a photograph a century older shares nothing with the phone's calibration, and pinning its principal point makes COLMAP tilt the camera instead.

dense

Key	Default
enabled	false	COLMAP's dense stereo over the colmap model; needs CUDA
max_image_size	2000	photographs are scaled down to this for stereo

The configs in the repository

File	What it is
projects/valencia/configs/cpu.yaml	the whole pipeline on the CPU, COLMAP computed here
projects/valencia/configs/gpu-dense.yaml	the same on a GPU, with COLMAP's dense cloud
projects/valencia/configs/no-colmap.yaml	the CPU run scored against a saved COLMAP model, for machines without COLMAP