Baselines

The paper compares its local optimizers with two families of established refiners, both run as ExternalRefit strategies (VisualGeometryToolkitCore's external refit contract: external_workspace, refine_external!) through the same refit layer (VisualGeometryToolkitEval.Evaluation) and under the same LO budget as every other arm: PoseLib's truncated refiners (in the package) and RansacLib's classical LO-MSAC (in the paper project accv2026/, see below). They are evaluation-only: they refine a selected seed at an absolute loss width, never inside the native RANSAC loop.

armstrategywhat runs
TLS-IRLSPoseLibLm(max_iter)PoseLib Levenberg–Marquardt on the truncated Sampson loss (TLSShape)
TLS-LZPoseLibLZ(max_iter)PoseLib's Le–Zach relaxed IRLS on the same truncated loss
TLS-IRLS-ŝPoseLibLmAdaptLossWidth()TLS-IRLS with a partition-anchored width, re-estimated between rounds (E only)
TLS-LOLebedaIterLO(opts, rng_seed) (accv2026)RansacLib's iterative least-squares ladder alone
TLS-LO⁺LebedaLOPlus(opts, rng_seed) (accv2026)RansacLib's full LO: least squares at the enlarged width, inner non-minimal resampling, per-sample ladder

PoseLib truncated refiners

PoseLibLm and PoseLibLZ refine relative pose (refine_relpose), fundamental matrices (refine_fundamental) and homographies (refine_homography) at the absolute loss width the refit hands refine_external! (a LossWidth). max_iter is their LM budget; the paper runs them at the evaluation's equal LO budget.

RansacLib LO (RansacLibLO)

accv2026/lib/RansacLibLO wraps Torsten Sattler's RansacLib LO-MSAC (Lebeda, Matas and Chum, BMVC 2012), run unmodified as a seeded local optimization for E, F and H.

  • Vendored byte-identical. RansacLib is vendored at upstream commit b952083 with its LICENSE and NOTICE (BSD-3-Clause), and declared in VENDORED.toml as upstream code outside the stack's naming rules. Vendoring removes the network dependency from the reproduction path.
  • The only code added is what RansacLib's template API asks of a user — the E/F/H solver plug-ins and their Sampson residuals — and a subclass that exposes the protected LocalOptimization / LeastSquaresFit entries so the LO runs from a caller-supplied seed.
  • Configuration is RansacLib's public LORansacOptions subset (LebedaLOOptions). PRESET_REFERENCE is RansacLib's own defaults; the paper builds budget-parity options from its protocol of record (accv2026/configs/protocol.toml, [arms.lebeda]). Both the options and the rng_seed are required: they are protocol values of the caller.
  • Deterministic. RansacLib's per-call mt19937(rng_seed) is its only source of randomness, so identical inputs give identical outputs. The wrapper also reports the number of linear solves per call, the budget-parity audit.
  • Part of the paper project, not the package. RansacLibLO builds a C++ shim locally and is not registered, so the registered LocalOptimizationDoneRight does not depend on it. The paper project devs it by path ([sources] in accv2026/Project.toml), and the two arms (accv2026/src/lebeda_lo.jl) implement VisualGeometryToolkitCore's external refit contract (external_workspace, refine_external!), like any third-party strategy.

Building it compiles a small C++ shim and needs git and a C++17 g++; Eigen comes from Eigen_jll (set EIGEN_INCLUDE to use another copy). From accv2026/:

using Pkg; Pkg.instantiate(); Pkg.build("RansacLibLO")

Loss width and frames

The refit's loss width is the absolute, unsquared truncation width in the pool frame — the same currency the PoseLib arms consume. RansacLibLO takes the squared width in the frame of its input points, so the wrapper converts once:

  • F and H — the pool frame is the model frame: pool-frame points, seed and width go through unconverted. On H the shim scores the codimension-2 Sampson error of the transfer constraint, whereas PoseLib's truncated H path truncates the symmetric transfer error: the same width, a different residual.
  • E — RansacLib's essential solver projects onto the essential manifold, which is valid only for calibrated coordinates. The wrapper feeds it calibrated points on the z = 1 plane, converted once from the pool frame by the stored pool-frame calibration Kc, and converts the width with exactly PoseLib's camera-aware rule, ½(1/f₁ + 1/f₂), so TLS-IRLS, TLS-LZ and TLS-LO⁺ consume the same pool-frame width.

This E arm is a declared pinhole exception. Calibrated geometry in the stack is otherwise on unit bearings from the camera's back-projection, never the z = 1 plane; RansacLib is third-party code run unmodified, so its essential solver's plane-point convention is accepted here, pinhole-only, and declared where it lives.

Citing

If you use the RansacLib arms, cite RansacLib and Lebeda et al.:

@misc{Sattler2019Github,
  title  = {{RansacLib - A Template-based *SAC Implementation}},
  author = {Torsten Sattler and others},
  URL    = {https://github.com/tsattler/RansacLib},
  year   = {2019}
}
@inproceedings{Lebeda2012BMVC,
  title     = {{Fixing the Locally Optimized RANSAC}},
  author    = {Karel Lebeda and Jiri Matas and Ondrej Chum},
  booktitle = {British Machine Vision Conference (BMVC)},
  year      = {2012}
}

One disclosed difference from Lebeda's original code: its iterated least squares used Sampson-gradient reweighted fits, whereas RansacLib's refits are plain least squares, used as-is.