Overview
LocalOptimizationDoneRight holds the local optimizers (LO) of the ACCV 2026 paper Robust Local Optimization Done Right: the strategies the paper proposes, the alternatives it studies, and the research variants behind its ablations. They implement VisualGeometryToolkitCore's local-optimization contract, so each one composes with any scorer in VisualGeometryToolkitEstimation's RANSAC. The repository has two halves:
the package (src/, ext/) | the strategies, the joint-(θ,z) lift core and its continuation schedules, the cross-validated operating points, the strategies' TOML registry, and PoseLib's truncated refiners (a package extension) |
the research project (accv2026/) | the paper's arm rosters, configs, drivers, renderers, the RansacLib baseline and the manuscript; see Reproducing the paper |
The package depends on VisualGeometryToolkitCore (the contracts, the losses, the width axes) and VisualGeometryToolkitEstimation (the problems, the RANSAC loop, the core baselines nothing (no local optimization) and Irls{K}). PoseLib is a weak dependency.
Scoring versus refinement
RANSAC asks two different questions of a robust objective. Scoring ranks minimal-sample hypotheses and must be sharp: a single wrong inlier set loses the pair. Refinement moves the selected hypothesis toward the solution and needs a soft objective: gradient from correspondences that the seed currently rejects. The stack keeps the two roles behind separate contracts, so a scorer and a local optimizer are chosen independently. With the camera and the pixel correspondences cs of the Getting started scene:
using VisualGeometryToolkitEstimation, LocalOptimizationDoneRightusing VisualGeometryToolkitCore: WelschShape, StudentTUniformScoring, JeffreysPrior, FlatLabelPrior# the core default: MSAC at the class's threshold + TLS-IRLS following itresult = find_essential(camera, camera, cs)# the paper's recipe: Student-t + uniform selection (outlier half-width 250 px),# Welsch-HQ refinementresult = find_essential(camera, camera, cs; scoring = StudentTUniformScoring(EssentialProblem, 250.0; scale_prior = JeffreysPrior(), label_prior = FlatLabelPrior()), local_optimization_strategy = HQ{WelschShape}(; efficiency_constant = WELSCH_HQ_EFFICIENCY_CONSTANT_ESSENTIAL))Strategies
| strategy | path | what it does |
|---|---|---|
HQ{K} | :hq | joint-(θ,z) half-quadratic lift on loss K; the paper's method (Welsch-HQ) |
WelschHQNewton, WelschHQEcm, BarronHQ, GmHqEps | :hq | research variants of the lift |
Gnc{Irls{K}} | :gnc | graduated non-convexity over a core IRLS |
StudentTUniform | :irls | the Student-t + uniform mixture IRLS (EM), self-scaling |
StuTuBCD | :bcd | block-coordinate ascent of the scorer's own objective (the score-matched optimizer) |
StuTuHQTabulated, StuTuRegemm | :hq, :gnc | mixture lifts |
BarronAdaptive | :irls | Barron's adaptive loss under IRLS |
PoseLibLm, PoseLibLZ, PoseLibLmAdaptLossWidth | :irls, :gnc | PoseLib's truncated refiners (external refiners) |
The lifts run on VisualGeometryToolkitCore's chart-rows contract, which VisualGeometryToolkitEstimation implements per problem, so one code path serves every liftable problem; see Lift core and schedules.
Widths
A width-carrying strategy (HQ, Gnc, WelschHQNewton, BarronHQ, GmHqEps, BarronAdaptive, and the core Irls) holds its width mode as a type parameter, resolved once per call by VisualGeometryToolkitCore's loss_at:
efficiency_constant = m(EfficiencyConstant(m)): the loss width ism·ŝ, withŝthe scale of the consensus set the scorer selected (the paper's partition anchoring).mis always the plain multiple of the scale, whatever the loss shape;loss_width = w(LossWidth(w)): an absolute width, in the caller's units;- neither: the width follows the scorer. The loss's decision point is placed at the scorer's
decision_boundary; a scorer without one is refused at entry.
Give at most one of efficiency_constant and loss_width. The cross-validated constants of Operating points and TOML are efficiency constants, tuned per model class.
with_width(lo, w) returns the same strategy at another width mode. The self-scaling strategies (StudentTUniform, StuTuBCD, WelschHQEcm, the mixture lifts) carry no width; they start from the consensus scale.
External refiners
PoseLib's truncated refiners are ExternalRefit strategies: they refine a seed at an absolute loss width through VisualGeometryToolkitCore's external-refit contract (external_workspace, prepare_external!, refine_external!), outside the RANSAC loop. The PoseLib package extension implements those methods. The paper project's RansacLib arms implement the same contract; see Baselines.
Operating points and TOML
tuning_constant(problem, shape; features, dataset) returns the cross-validated efficiency constant of a loss on a model class, with its provenance (the accv2026 config that produced it); the WELSCH_HQ_EFFICIENCY_CONSTANT_* constants are its Welsch-HQ entries.
LODR_STRATEGY_TAGS registers the strategies' TOML tags. Merge it with the core's to read a strategy table of either package, such as a find_* pipeline's local_optimization_strategy = { type = "hq", loss = "welsch", … }:
lo = optimization_strategy_from_toml(d; class = :essential, tags = merge(VisualGeometryToolkitEstimation.OPTIMIZATION_STRATEGY_TAGS, LocalOptimizationDoneRight.LODR_STRATEGY_TAGS))In such a table, efficiency_constant = "cross_validated" resolves through tuning_constant.
Evaluation
The paper's evaluation engine (pair geometry, the one refit layer, post-LO error matrices, the sample-budget engine, live RANSAC runs, the Roster of arms and its gate-enforced readers, the selection stream) is VisualGeometryToolkitEval.Evaluation, written against the core contracts alone. It evaluates any strategy here without per-type code: a native strategy through locally_optimize!, an external one through refine_external!.
Defaults of record
Every strategy default lives in configs/default.toml, read through one function, _config_default(table, key), which throws on a missing table or key; there is no second set of Julia literals. The file is parsed at precompile time, so an edit forces a recompile.
| table | holds |
|---|---|
[optimization_strategy.<tag>] | each strategy's defaults (max_iter, damping, schedule), keyed by its LODR_STRATEGY_TAGS tag |
[optimization_strategy.poselib] | cheirality_vote_points: points in the cheirality vote of an E decomposition |
[optimization_strategy.poselib_adapt] | PoseLibLmAdaptLossWidth's total budget, rounds, min_support and operating efficiency constant |
[[tuning_constants]] | the cross-validated operating points with their provenance |
A paper with its own protocol states its values explicitly; the ACCV 2026 protocol of record is accv2026/configs/protocol.toml.
Frames
Frame in, frame out: public functions return results in the frame their inputs came in. The one declared exception is the benchmark pool, stored in NormalizedImageFrame; every API where pool data enters or leaves names that frame. Calibrated geometry is on unit bearings, except for the vendored RansacLib essential-matrix arm, a declared third-party pinhole exception (see Baselines).