Source: https://github.com/cdcseacave/openMVS_sample
Revision: 8b67734532c4391604c646071938e7ac8ae9486d (v3.0.0).
The upstream repository supplies AGPL-3.0; its complete LICENSE is retained in
LICENSE-OpenMVS-sample.txt. Original photographs and reconstruction outputs
remain attributed to their original source. Consult upstream for material-specific
rights. This notice does not change the license of the surrounding website.
These files are converted upstream reconstruction data, not a reconstruction executed for the article. Changes made for display:
x y z r g b (24 bytes/vertex).
Faces: little-endian uint32 triplets. Coordinates have relative, not metric, scale.scene_dense.jpg and scene_dense_mesh_refine.jpg copied unmodified upstream
preview images. They are not rendered screenshots from this website.features.json: independent OpenCV SIFT matching on the resized photo pair;
these are not upstream SfM feature tracks. Includes actual inliers and outliers.bundle.json: independent synthetic 4-camera / 18-point least-squares example;
these are not Castle observations, poses, or BA history. Focal length and the
first two camera poses are fixed. Each budget starts from the same initialization.plane-sweep.json, sweep-0.webp, sweep-1.webp: independent real-photo
calibrated two-view plane sweep. Uses the final intrinsics/distortion printed
in the pinned upstream log, scales calibration for the thumbnails, undistorts
them, and independently estimates relative pose. Translation has unit norm;
depths are baseline-relative reference-camera Z, not metric or the full model’s
display coordinates. 161 inverse-depth candidates, fixed fronto-parallel normal,
11/21-pixel windows. ZNCC and normalized mean SSD use uncompressed grayscale
patches and bilinear warping. The affine-exposure condition is a controlled
1.2 gain +20 offset without clipping; ZNCC is mathematically invariant to it.
This experiment performs no PatchMatch, multi-view aggregation, or dense fusion.Regenerate with python scripts/prepare_reconstruction.py from the repository
root. Requires numpy, scipy, opencv-python and Pillow. Pinned source downloads
are cached in tmp/mvs-source/; hashes and original counts are in manifest.json.