Repository navigation
Releases
Speed optimizations.
1.2-2x faster than 0.3.0, same output
- GIL released during estimation; thread-safety fixes in the degeneracy RNG and table init
- Faster fixed-size solvers and degeneracy recovery, fewer allocations, pooled LO hash nodes
- AVX kernels picked at run time on x86, bit-identical to the scalar path (
PYDEGENSAC_SIMD=0forces scalar) - Windows wheels use thread-safe OpenBLAS, so concurrent calls are safe
- Examples switched from AKAZE to SIFT so they run on OpenCV 5

Fix MacOS build, speed-up everything
Highlights
macOS builds were broken and are now fixed. Every macOS wheel ever published
skipped all of its LAPACK calls: the C sources guarded them behind
#ifdef _WIN32 / #ifdef __linux__, and macOS defines neither. Local
optimisation — the LO in LO-RANSAC — never ran. Results were degraded, not
broken, so nothing failed loudly: it cost 0.11 mAA on fundamental matrix and
0.21 on homography on public data. If you have ever benchmarked pydegensac on
a Mac, that number was wrong. Linux and Windows were unaffected.
1.7–2.9x faster, at unchanged accuracy: 2.2x (F) / 1.7x (H) on Linux,
2.3x / 2.9x on Apple Silicon. From removing a per-iteration RNG re-key, 45
redundant passes in the covariance accumulation, a division per correspondence
in inlier scoring, and seven of eight divisions in the homography error
function — plus LTO.
cp314 wheels for Linux, macOS and Windows.
Notes
- Results change relative to 0.2.2. On macOS substantially (the fix above);
elsewhere in the last bits, from deliberate floating-point reassociation. Every
such change was gated over 1000 seeds × 10 image pairs against inlier counts,
ground-truth error and model quality — no distributional difference (min
p = 1). Fixed-seed output is reproducible as before, but not bit-identical to
0.2.2. - Linux wheels still target glibc 2.17, so CentOS 7 and Ubuntu 18.04 keep
working. They are also faster than 0.2.2's for reasons unrelated to the
optimisation work — the build now prefers OpenBLAS over the 2012 reference
LAPACK it used to vendor. ⚠️ Do not usepydegensac==0.1.2with numpy 2.x — that combination
silently returns every correspondence as an inlier. It was never yanked, so
an old pin can still resolve to it.
Benchmarks
New benchmarks/ suite: accuracy-vs-compute against cv2.RANSAC,
cv2.USAC_MAGSAC and poselib on public data. pydegensac is now statistically
tied with the best available on fundamental matrix, and the second-cheapest
homography estimator in the roster. Full results and protocol in
docs/reports/2026-08-11-ransac-benchmark.md.
Expose random seed parameter to make it deterministic
Merge pull request #38 from ducha-aiki/speed-up Expose seed and make LAF allocation to be conditional
Works on more modern python
Merge pull request #31 from ducha-aiki/fix-logic fix wrong behaivour introduced by codex
Works in PIP
v_0.12 added info about pip installation
Wheel release
First release
v_0.1 Initial commit