Skip to content

Releases

This page records pyTheia releases. It no longer mirrors the historical TheiaSfM upstream changelog; older version notes referred to upstream tarballs and features that this fork no longer ships.

1.1.0 (current)

Bulk calibrated two-view estimation now carries monodepth depth priors and returns relative depth-map scale.

Bulk two-view API

  • BulkEstimateTwoViewInfo accepts optional per-correspondence depth_i / depth_j (1-D float arrays, length \(N\)) and sets Feature.depth_prior when building flat correspondences. With EstimateTwoViewInfoOptions.use_monodepth=True and ≥95% valid depths, each pair uses the calibrated monodepth 3-pt path (same as scalar EstimateTwoViewInfo).
  • Return dict gains scales: per-pair TwoViewInfo.scale_estimate (\(-1\) when unset / unused). Existing keys (success, orientations, positions, inlier_offsets, inlier_indices) are unchanged.
  • Smoke test: pytests/sfm/test_bulk_monodepth_two_view.py.

Narrative: RANSAC — monocular-depth-assisted two-view, Performance — GIL / bulk, Estimators — relative pose.

1.0.9

Faster two-view / PnP RANSAC paths: PoseLib-adapted minimal solvers plus dense LM local optimization that replaces Ceres in the LO hot path.

Two-view estimation

  • Sturm-sequence 5-point essential-matrix solver (FivePointRelativePoseSturm, theia/math/sturm.h), adapted from PoseLib / Nistér. Default for EstimateRelativePose / essential-matrix RANSAC via RansacParameters.use_sturm_5pt (default true).
  • RANSAC-loop speedups for relative pose: Sampson-only scoring with a single final essential-matrix decomposition, vectorized Sampson residuals, less allocation churn in SampleConsensusEstimator.
  • Monodepth 3-point relative-pose solvers and estimators (relative_pose_monodepth_3pt, EstimateMonoDepthRelativePose + shared-/varying-focal variants), adapted from PoseLib RePoseD DingRePoseD2025. Wired into EstimateTwoViewInfo via use_monodepth / monodepth_shared_focal when correspondences carry Feature.depth_prior.

Narrative: RANSAC, Performance, Pose — Sturm 5-pt, Pose — monodepth 3-pt, Pose — dense LM refiners, Math — Sturm, Math — lmlsq.

Dense LM RANSAC local optimization (use_lo)

New tiny Levenberg–Marquardt stack under theia/math/lmlsq/ (truncated loss + dense normal equations). Model-specific refiners (PoseLib-style Jacobians, Theia types) replace Ceres problem setup on every LO call:

Estimator Refiner Residual / DoF
EstimateRelativePose refine_relative_pose Sampson; 5
EstimateMonoDepthRelativePose (+ shared/varying focal) refine_monodepth_relative_pose Sampson (+ \(F\)) and depth reprojection; 7–9
EstimateCalibratedAbsolutePose refine_absolute_pose Reprojection; 6

Full-scene / multi-view bundle adjustment is unchanged and still uses Ceres. Details: RANSAC — local optimization, Pose — dense LM refiners.

Docs and licensing

1.0.x

Stable Python packaging line with MkDocs documentation, cross-run alignment APIs, and continued binding coverage for cameras, BA, pose solvers, view graphs, and I/O (COLMAP, Bundler, NVM, PLY, Nerfstudio, SDFStudio, etc.).

Highlights since 0.6.0

  • Versioning aligned across root VERSION, CMake project(), and pyproject.toml.
  • Documentation maintained as MkDocs under docs/content/ (Sphinx-era pages removed).
  • Cross-run alignment APIs documented in Cross-run alignment and Transformations (BA relative pose edges, Sim(3) pose graph, scale-invariant odometry).
  • Examples: lightweight scripts in pyexamples/; showcase examples/vismatch_sfm/; Nerfstudio export via pyexamples/nerfstudio_export_reconstruction.py.
  • Wheels: manylinux builds for Python 3.8–3.13 via Docker (urbste/pytheia_base).
  • 1.0.9: Sturm 5-pt + monodepth 3-pt solvers; dense LM RANSAC LO for relative / monodepth / calibrated absolute pose — see 1.0.9 notes.
  • 1.1.0: BulkEstimateTwoViewInfo depth priors + scales — see 1.1.0 notes.

Known limitations

  • C++ unit tests exist but are mostly unwired in CMake; CI builds wheels without running pytest or CTest. A growing subset of pose / estimator / lmlsq tests can be enabled with BUILD_TESTING=ON.
  • Feature detection / descriptor matching in C++ is deprecated in favor of Python-side pipelines; see 0.6.0 notes below.

0.6.0

Major cleanup: smaller surface area, Python-first imaging and features, fewer vendored or optional dependencies in the default build path.

Removed or deprecated in this line

  • Raster / EXIF in C++: The theia/image stack and OpenImageIO integration are gone. Images and EXIF are expected in Python (OpenCV, Pillow, etc.). C++ keeps geometry, reconstructions, cameras, BA, and solvers.
  • Feature and descriptor I/O: Reading/writing bundled keypoint–descriptor files and SIFT key formats has been removed from the library and from Python bindings. Feature detection and matching stay outside the core package; you pass correspondences and use Feature, FeatureCorrespondence, and estimators as needed.
  • Python API trim: Descriptor-centric matching helpers (e.g. KeypointsAndDescriptors, brute-force / cascade hash matchers, Fisher-vector globals, two-view verification types that depended on that stack) are no longer exposed on pytheia.matching / pytheia.sfm. FeaturesAndMatchesDatabase, InMemoryFeaturesAndMatchesDatabase, ReconstructionBuilder, and ReconstructionBuilderOptions are also removed from Python (C++ Theia still uses the match database inside its own reconstruction-builder path).
  • Exports and extras: Legacy PMVS export and other dead or redundant tooling paths have been removed or stripped from docs and CMake where applicable.
  • Documentation: Sphinx-era pages were replaced with MkDocs; narrative docs focus on what pyTheia actually exposes and how to wire Python-side features into SfM.

Cross-run alignment additions (0.6.x)

New capabilities for aligning two independent reconstructions (segment + run). Narrative docs: Cross-run alignment.

  • BundleAdjustReconstructionWithRelativePoseEdges — full BA plus SE(3) pose-to-pose edges that snapshot run odometry at setup time.
  • RelativePoseConstraint.scale_invariant_translation — optional scale-invariant odometry (translation direction + optional log-magnitude) via ScaledRelativePoseError.
  • BundleAdjustReconstructionWithConstantTracks — control-point BA: fixed map tracks, optimized run cameras.
  • Sim(3) view helpers — GetSim3LieFromView, RelativeSim3BetweenViews, SetViewCameraFromSim3Lie.
  • Cross-reconstruction Sim(3) pose graph — AlignReconstructionsWithPoseGraph, CrossReconstructionSim3PoseGraphOptimizer, sequential / anchor / scale-smoothness edges (CrossViewAnchorEdge, SequentialSim3Edge, …). Documented under Transformations → pose graph.

Still here

  • Core SfM pipelines, bundle adjustment, cameras, pose / triangulation solvers, view graph, I/O for reconstructions (binary/JSON, Bundler, NVM, COLMAP text, PLY, Nerfstudio, SDFStudio, etc.), and pybind11 Python API for the remaining types and functions.

If you need the old upstream release notes for academic comparison, see the TheiaSfM repository tags and history.