pyarchinit-photogrammetry-core

Native kernels of pyarchinit-photogrammetry, written in Rust (PyO3 + rayon) and imported as ppg_core. The package is optional: pyarchinit-photogrammetry keeps the numpy version of every kernel, uses it when ppg_core is missing or when PPG_NATIVE=0, and its tests compare the two.

Kernels: fusion of the depth maps of one photo, depth z-buffer and normals, projection of points with the COLMAP camera models, scoring of the photos of an orthomosaic tile, agreement of points with a depth map (confidence), PatchMatch multi-view stereo of one photo (patchmatch_depth: a plane per pixel, normalised cross correlation with the best of its neighbouring photos, red-black propagation and random refinement, from random planes or from a depth map), the GPU kernels (wgpu: SIFT, descriptor matching, PatchMatch, fusion), RTKLIB's post-processing of GNSS observations (ppk_solve, from 0.5: what rnx2rtkp -k options -o solution files does, with RTKLIB built into the extension), and the decoding of Hatanaka-compressed RINEX (crx2rnx, from 0.6: Compact RINEX 1.0, 3.0 and 3.1 into the RINEX that RNXCMP's CRX2RNX writes, by a decoder of our own written from the description of the format in Hatanaka (2008), Bulletin of the Geographical Survey Institute 55, 21-30). From 0.7 (API 10) the kernels of the TIN (engine.tin): the constrained Delaunay triangulation of the vertices with the break lines as constraints (tin_triangulate, with the crate spade, robust predicates in double precision; constraints that cross are split at a new vertex), the TIN rasterised on a grid (tin_rasterize), its contour segments (tin_contours) and its exact cut and fill against a plane (tin_volume).

Build and install in the environment of pyarchinit-photogrammetry (a Rust toolchain is needed):

pip install ./native        # or: uv pip install ./native

Building it compiles RTKLIB's C sources too (vendor/rtklib, with the cc crate): a C compiler is needed (Xcode's command line tools, gcc or clang, MSVC on Windows). RTKLIB is built as C99 and without fused multiply-add (-ffp-contract=off, /fp:precise with MSVC): the integer ambiguities of a kinematic solution hang on the last bit of the arithmetic, and the arm64 compilers, which fuse by default, would otherwise fix other epochs than the x86 builds of the same project (on RTKLIB's test data 115 of 120 instead of about 40).

License: GPL-2.0-or-later. The extension includes RTKLIB-EX v2.5.1 (https://github.com/rtklibexplorer/RTKLIB, commit 62d4677ed8425a4e2748c6d390b500d1afb493fc), Copyright (c) 2007-2020, T. Takasu, distributed under the BSD 2-clause license: its text is in vendor/rtklib/license.txt and in NOTICE, which come with the package. RTKLIB's optional routines of the IAU SOFA library (sofa.c, a licence of their own) are not included.

The Rust crates built into the extension are under licences compatible with the GPL (MIT or Apache-2.0: PyO3, numpy, rayon, wgpu, spade and their dependencies).