[servantftransperfect]

Description

Several nodes gave different results from one run to the next on the same inputs. The main cause was random generators shared by all the threads of a parallel loop (a data race, and RANSAC results that depend on thread scheduling). Other causes were results collected in thread completion order, sorts without a tie-break, and files read in directory order.

This PR makes these nodes repeatable for a given thread count, and in most cases independent of it.

Features list

• [numeric] Add createSeededGenerator(baseSeed, ids...): a generator for one work item (image pair, track), seeded from a base seed drawn once outside the parallel loop.
• FeatureExtraction (sift, akaze): features are kept in keypoint order instead of thread completion order. dspsift already was. Adds a features_sift test comparing 1 and 4 threads.
• FeatureMatching, GeometricFilterEstimating: one generator per image pair in robustModelEstimation.
• RelativePoseEstimating: one generator per pair (the result no longer depends on thread count or chunking). Pairs are processed in bounded batches and written in pair order; the output format is unchanged.
• SfmBootstrapping, NodalSfM: pairs are sorted after loading, and best pair / best view picks break ties on the view ids.
• GlobalRotationEstimating, PanoramaRefining: pairs are sorted after loading (directory order is unspecified).
• PanoramaSeams: tiles are ordered by their number, so imageCounter and the new view ids are repeatable.
• StructureFromMotion, SfMTriangulation: one generator per track in the multi-view triangulation; next best views break ties on the view id.
• SfmExpanding: one generator per track in the triangulation; both expansion policies break ties on the view id; resections are applied in view id order (with rigs, the last finished thread used to set the shared pose).
• GlobalSfM: one generator per pose pair for the relative rotations, sorted before the rotation averaging.
• LdrToHdrSampling, LdrToHdrCalibration: fixed seed instead of std::random_device.

Implementation remarks

• RANSAC now draws different numbers than before, so results are not bit-identical to previous versions. They were not repeatable before either.
• Still not repeatable:
  • Meshing / LidarMeshing: the default seed of 0 means std::random_device, the helper points noise uses a shared generator, and the graph weights are float sums accumulated from several threads.
  • GlobalSfM translation triplets: the edge coverage is greedy across threads (and reads the shared set without a lock), so which triplets are solved depends on scheduling.
  • KeyframeSelection: writes a random Exif:ImageUniqueID on purpose.
• Across machines with different core counts, multithreaded Ceres (bundle adjustment, rotation and position averaging) may still differ in the last bits. Using the same maxCores on every machine avoids it.

Tests

• All 109 ctest tests pass (OpenMP enabled), including the new features_sift.
• Not done yet: two-run output comparisons on a real dataset with different --maxCores.

🤖 Generated with Claude Code