Reproducing the paper
The research project accv2026/ produces every figure and table of the paper. It is a DrWatson project that devs the package by path ([sources] in accv2026/Project.toml). The package, its tests and docs, and accv2026/ are one Pkg workspace: they resolve together, in one Manifest.toml at the repository root that is not committed. accv2026/REPRODUCE.md is the authoritative experiment → artifact index and records the data provenance; this page summarizes it.
The package and the research project split by one rule:
| Belongs in the package | Belongs in accv2026/ |
|---|---|
| the local optimizers a second paper reuses | paper drivers, figures, arm rosters |
| registered LO baselines (PoseLib) | the unregistered RansacLib baseline; anything that exists to make one comparison defensible |
Layout
| path | contents |
|---|---|
accv2026/src/ | the shared prelude: config.jl (entry), registry.jl (every path), arm_rosters.jl (THE arm table and the paper's Roster), protocol_folds.jl (the run-key fold recipe), cell_stages.jl (the shared stage builders), figure_style.jl, … |
accv2026/scripts/ | one executable driver per experiment: fig*, supp_*, the pool producers, the table drivers (timing_table.jl, e2e_ransac.jl), the SupeRANSAC cell superansac_e2e.jl, polish.jl, heb_inlier_stats.jl |
accv2026/scripts/tools/ | the CLIs the drivers launch: run_sequential_cv.jl (the tuning), run_polish.jl, the renderers, the pool builders, make_facet_grid.py |
accv2026/scripts/reproduce/ | the launchers of VisualGeometryToolkitEval's runner on the stage graph of record, configs/reproduce.toml (reproduce_results_dgx.sh on Slurm, reproduce_results_desktop.sh locally; --dry-run, --only=…) |
accv2026/configs/ | the configs of record: one TOML per cell, and the protocols folded into every run key |
accv2026/data/ | two per-machine symlinks: exp_raw (immutable inputs) and exp_pro (add-only results), or VGTK_DATA_ROOT / VGTK_RESULTS_ROOT |
accv2026/lib/RansacLibLO/ | the classical RansacLib LO baseline (vendored C++ shim, Pkg.build("RansacLibLO")) |
accv2026/envs/ | auxiliary environments (SupeRANSAC) |
accv2026/paper/ | the manuscript; renders land in paper/figures/ |
accv2026/test/ | the project's tests (julia --project=. test/runtests.jl) |
accv2026/REPRODUCE.md | the experiment → artifact index and the data provenance |
The package itself is src/ (the strategies in src/strategies/, the joint-(θ,z) lift core and its continuation schedules in src/lift/, the cross-validated operating points in src/tuning_constants.jl, PoseLib's truncated refiners in src/poselib_arms.jl), ext/ (the PoseLib extension), configs/default.toml (the strategies' defaults of record and tuning constants), test/ and docs/.
Drivers and stages
Every driver is stage-idempotent: a stage whose output exists is skipped, --force re-runs, --stage=NAME runs one stage, and --plan prints the stage plan without running anything. From accv2026/:
julia --project=. scripts/fig3_composed_miscal.jl --planjulia --project=. scripts/fig3_composed_miscal.jl --stage=cvEach driver declares its data inputs (require_input, with provenance), tunes only through the single cross-validation path, audits every tuning artifact before consuming it (audit_operating_points, VisualGeometryToolkitEval's Evaluation), and ends by rendering its panel into paper/figures/. The whole paper is one command of the runner, bash accv2026/scripts/reproduce/reproduce_results_dgx.sh (or _desktop.sh); its --dry-run lists every stage and every missing input first.
Config-driven cells and keyed output directories
A cell is a config of record (configs/<cell>.toml) plus the driver that runs it. Its effective config — the TOML with the protocols of record folded in by src/protocol_folds.jl — is hashed into its run key (VisualGeometryToolkitEval Experiments.run_key, the SHA-256 of the sorted TOML; the [render] section is excluded). Results are written to keyed directories such as phototourism/sift/E/Miscalibration/sequential-cv-<key>/ under the add-only results root, never rewritten. Consumers re-derive a producer's key from its config name, so keys are derived, never copied.
| protocol of record | governs |
|---|---|
configs/tuning.toml | every [tuning.<name>] entry: windows, spacings, expand-on-boundary rules, objective, budget sampling, cross-validation |
configs/cv_protocol.toml | tuning at ζ = 1 and the per-dataset cross-validation splits |
configs/miscal_protocol.toml | the ζ evaluation ladder |
configs/sample_budget_protocol.toml | per-dataset stream depth and checkpoint schedule |
configs/protocol.toml | the experiment protocol: the equal LO budget, the Lebeda ladder, every per-arm deviation, and the result-changing values of each experiment |
The sequential cross-validation
Every two-view cell is tuned by the sequential CV, scripts/tools/run_sequential_cv.jl: two one-dimensional cross-validations in sequence rather than one product grid.
- Scorer — the selector's width
a*on the selection-only error →scorer_cv.jld2. - Refiners — each arm's width at that
a*, read back from the scorer artifact →refiner_cv.jld2. ASelectionWidtharm (its refit width is the selection widtha*) has nothing to tune here. - Evaluation — the ζ miscalibration ladder on the test scenes at the tuned points →
operating_points.jld2, the sidecar every downstream cell reads.
Each stage expands its window when the argmin lands on its boundary; a surviving clip is recorded per arm and the readers refuse it.
Figure → config → driver
The paper's figures stack per-dataset panels into one grid. The figure configs configs/fig3_miscal_grid.toml and configs/fig4_budget_grid.toml name each row's panel and the driver stage that produces it; they compose and run no experiment.
| paper | rows (dataset · model) | producing driver (stage) | cell config |
|---|---|---|---|
| Fig. 1, perturbed-seed basins | PhotoTourism · E | fig2_basin.jl (basin_cv, ladder_test, render) | pt_e_basin.toml |
Fig. 2, scale miscalibration — fig3_miscal_grid.jl (compose) | PhotoTourism · E | fig3_composed_miscal.jl (cv, oracle, render) | pt_e_miscal.toml |
| ETH3D · F | supp_ETH_F_miscal.jl (render_main) | accv2026_eth_f_miscal.toml | |
| HEB · H | supp_H_miscal.jl (render_main) | heb_h_miscal.toml | |
Fig. 3, hypothesis budget — fig4_budget_grid.jl (compose) | PhotoTourism · E | fig4_budget.jl (render) | pt_e_budget.toml |
| ETH3D · F | supp_ETH_F_budget.jl (render) | accv2026_eth_f_budget.toml | |
| HEB · H | supp_H_budget.jl (render) | heb_h_budget.toml | |
| Fig. 4, least-squares polish | PhotoTourism · E | polish.jl (test, summary, render) | polish_pt_e.toml |
| Table 3, end-to-end RANSAC | PhotoTourism · E | e2e_ransac.jl (live, summary, table); SupeRANSAC's rows superansac_e2e.jl (inputs, cv, eval) | pt_e_e2e_table.toml, pt_e_superansac.toml |
fig3_composed_miscal.jl --stage=cv writes THE deployed operating-points sidecar of PhotoTourism E that the downstream PhotoTourism cells read, so it runs first. The supplement's figures and tables (the scoring-only miscalibration, the partition-anchoring ablation, the polish) and Table 1's timing are indexed in accv2026/REPRODUCE.md.
SupeRANSAC is an outside dependency: superansac_e2e.jl runs it in its own environment (accv2026/envs/superansac, PythonCall/CondaPkg with a source build of pysuperansac), detects whether it is installed and says how to install it. Without it, Table 3 is built from the stored SupeRANSAC results the cell's config names.
Data
Nothing large is committed. Two-view scoring pools are built by pool-producer cells (scripts/pt_e_pools.jl, scripts/heb_h_pools.jl, scripts/scoring_pools.jl --config=<name>) from the public datasets — IMC PhotoTourism, ETH3D, HEB — via scripts/tools/generate_scoring_pools.jl. The PhotoTourism E pool alone is about 2 TB. The data roots resolve through VGTK_DATA_ROOT / VGTK_RESULTS_ROOT or the data/exp_raw / data/exp_pro symlinks (accv2026/data/README.md). The cross-validation stages need a CUDA GPU. What the paper shows is tracked as the rendered .tex figures and tables under accv2026/paper/figures/; the result artifacts themselves are not yet published. With the produced artifacts synced, every --stage=render regenerates its committed panel on the CPU.
Rebuild the manuscript after rendering:
cd accv2026/paper && makeTests
The package's tests run from the repository root; the research project has its own suite, which checks among other things that REPRODUCE.md lists every driver:
julia --project=. -e 'using Pkg; Pkg.test()'cd accv2026 && julia --project=. test/runtests.jl