Developer notes

VisualGeometryToolkitEval is developed in the VisualGeometryToolkit.jl repository, a Julia workspace that holds the stack's packages, one directory each. The repository's root CLAUDE.md describes the workspace, and the conventions every stack package follows are in VisualGeometryToolkitCore's conventions/ folder.

Running the tests

From the repository root:

julia --project=. -e 'using Pkg; Pkg.instantiate()'          # once per clone
julia --project=VisualGeometryToolkitEval -e 'using Pkg; Pkg.test()'
julia --project=VisualGeometryToolkitEval -e 'using Pkg; Pkg.test(; test_args = ["pool_io|tuning"])'

The suite uses ReTestItems. Every test/*_tests.jl file holds one test item named after its file, and the items run in parallel on worker processes. The test_args filter items by name; check the reported total, because a pattern that matches nothing succeeds with zero tests. Fixtures that several items share are @testsetup modules in test/*_testsetup.jl.

The default suite includes the quality items: Aqua, ExplicitImports, a JET ratchet at zero reports, the calling conventions, a scan for undefined globals and the stack's naming check. The readme item runs every julia block of the README.

Building the documentation

julia --project=VisualGeometryToolkitEval/docs --startup-file=no VisualGeometryToolkitEval/docs/make.jl

The site is written to VisualGeometryToolkitEval/docs/build/.

Dependency boundaries

The package depends on VisualGeometryToolkitCore, VisualGeometryToolkitIO and VisualGeometryToolkitEstimation only. The evaluator is written against VisualGeometryToolkitCore's contracts, so RANSACScoringDoneRight, LocalOptimizationDoneRight and PoseLib never appear among its dependencies; the projects that evaluate them depend on this package instead.

Four source files are free of dependencies by contract, so that an environment that cannot load this package can include them by path: src/render/axis_style.jl, src/evaluation/cv_probe.jl, src/experiments/stages.jl and src/reproduce/reproduce_index.jl.

Frames

A scoring pool (ScoringPool) is stored in NormalizedImageFrame and says so (frame); a perspective-n-point (PnP) pool is stored in ImageFrame. The pool and residual files hold plain data together with the stored maps (source_map, destination_map), and a value leaves the pool frame only through those maps. Nothing in this package converts a frame. The raw readers of Datasets return pixels.