Conventions

The stack's shared conventions — naming, signatures, data arguments, frames and normalization, pixel origins, derivatives — are one canonical page in VisualGeometryCore's documentation, Conventions. This page does not restate them. It states the three things that are true of RansacScoringDoneRight and of nowhere else in the stack.

The pool frame is the one declared exception to "frame in = frame out"

Everywhere else the rule is pixels in, pixels out, with normalization private to one estimator. The benchmark pool is the exception, and it is declared rather than silent. The pool record — VisualGeometryCore's PoolPair (its Pools module documents the record, its maps and its accessors) — stores its correspondences and models in NormalizedImageFrame: image centre, scaled by HALF the image diagonal, so the image corners lie at unit radius. The record's residual matrix is not part of it: a pool's residuals are a separate num_points × num_models Float32 matrix in the same unit, and every pool scorer here takes the record AND its matrix — best_pool_model(scoring, pair, R, …), pool_model_score, precompute_pair(pair, R), stage_ig_shared_from_pairs(pairs, residuals). No scorer opens a file.

The exception exists because the pool is shared STORAGE, not an estimator's input: a pool is written once and scored many times, across scenes whose images differ in size. Storing one frame keeps a single global outlier half-width and one IG scale prior correctly specified across the whole benchmark; converting per pair at scoring time would not.

What the rule then demands of this package:

• Every width on pool data is in the pool frame's unit and says so: the outlier half-width, the loss width, the IG scale prior, every grid_evaluate grid and POOL_FRAME_R_MAX.
• There is one exit to pixels. pixel_residuals is the image-plane Sampson residual of the pool's models in pixels, computed from the light record alone: the correspondences come back through the stored maps (VisualGeometryCore's pixel_csponds), each model by composition with the normalization's sides (inv(D) ∘ M ∘ S). It never reads the stored residual matrix, so it does not depend on that matrix's residual kind — tangent Sampson for an essential pool, image-plane Sampson for F and H. residual_unit_scale is a CONVENTION, not a conversion — the two images are scaled independently, so no single scalar relates the frames; use it for a soft initial guess and never for anything a bound gates.

simfitpp_select on a pool record is the deliberate exception to the exception: it is pixels in, pixels out, running on pixel_residuals so that SIMFIT++'s published pixel bounds (0.25–8 px) mean what they say.

POOL_FRAME_R_MAX = 1 is a declared constant, not a tuning knob

POOL_FRAME_R_MAX is the upper limit of the domain [0, r_max] over which the density normalizer Z̃ of every redescending marginal (Welsch, smooth-TLS, Geman-McClure) is integrated. It is in the pool frame's half-diagonal unit.

It reads like a geometric bound, and it was first justified as one — under the FULL-diagonal frame every point lies within radius ½ of the centre, so no residual exceeds 1. Pools are written in the HALF-diagonal frame, where a point-to-point residual reaches 2, so that justification no longer holds and the constant truncates the outlier support at half the geometric maximum.

It stays 1 deliberately. Every constant tuned for these marginals — the cross-validated prior centres and outlier half-widths, and the published tables built on them — was cross-validated with r_max = 1, so the value is part of their definition, and one value keeps the truncation identical across scenes, probe and evaluation. Changing it moves every marginal score and invalidates every constant tuned against it.

PredictiveMarginalScoring is an option, not a default

PredictiveMarginalScoring incorporates the model's estimation covariance Σ̃_θ into the per-point scores. It is kept, and it is not the default, because it was measured rather than assumed.

On the reference rerun — 320 tasks, thresholds tuned per scoring AND per stage, paired seeds — against model-certain scoring, pooled ΔAUC@10°:

Its one clear win is the essential matrix after local optimization at N = 60 (+0.016). It costs 13–82× per hypothesis at N = 60 and 68–565× at N = 250, even with the kinetic enumeration. The loss is largest where the predicted variance is most heterogeneous across inliers.

Two things that verdict does NOT say. The search is EXACT — the enumeration visits every top-k set attainable at any μ > 0 — so this is a property of the FIRST-ORDER predictive model, not of the implementation. And whether linearization is the cause is open: an oracle-covariance experiment settles it. The option stays for that reason.

See also

• Conventions — the shared text this page links to rather than restates
• API Reference — the functions named above