Run panoramic dense stereo from two ERPs¶

Level: Start/compose · Input: two canonical ERPs, with optional relative pose and baseline · Output: radial range, validity, confidence, cross-view disagreement, pose and diagnostics

PanoramicDenseStereo is the A-to-Z application. It does not treat a rectified pinhole matcher as if it understood an ERP. It resolves pose and scale, creates paired rectified views from spherical geometry, runs one dense matcher at a time, converts disparity to radial range, and fuses range in the reference ERP.

ERP A + ERP B
      |
      +-- use supplied R,t ---------+
      |                             |
      +-- spherical features -> relative pose
                                    |
                 baseline / explicit scale-free status
                                    |
              paired rectified gnomonic views
                                    |
          classical dense stereo, one pair at a time
                                    |
                 disparity -> radial range
                                    |
            backproject -> confidence-weighted fusion
                                    |
       ERP range + validity + confidence + disagreement

Shortest call with known pose¶

from panorai.applications import PanoramicDenseStereo

stereo = PanoramicDenseStereo(pose_mode="provided")
result = stereo(
    erp_a,
    erp_b,
    rotation_source_from_reference=R_b_from_a,
    translation_source_from_reference=t_b_from_a_m,
    units="m",
)

range_m = result.radial_range
valid = result.validity
confidence = result.confidence
matching_confidence = result.matching_confidence
consistency_confidence = result.consistency_confidence
disagreement_m = result.disagreement

The pose convention is X_source = R_source_from_reference @ X_reference + t_source_from_reference. The norm of supplied t is used as the baseline unless baseline= explicitly rescales its direction.

Let PanorAi estimate relative pose¶

stereo = PanoramicDenseStereo(pose_mode="auto")
result = stereo(erp_a, erp_b, baseline=baseline_m, units="m")

This route reuses SphericalFeaturePipeline.for_relative_pose() and SphericalRelativePoseEstimator. The estimator’s existing quality policy is an enforced gate by default; rejection reasons are not converted into a dense result. Two central images determine translation direction, not metric scale. If baseline is omitted, the result is labelled scale-free and its range unit is baseline-units; it must not be described as metres.

Exchange the pinhole matcher¶

from panorai.applications import ADCensusStereoDepth, PanoramicDenseStereo

stereo = PanoramicDenseStereo(
    matcher=ADCensusStereoDepth(),
    route="rectified-views",
)

The six compatible matchers are documented in non-learned rectified stereo. They operate only after spherical planning and projection have produced rectified pinhole pairs.

Choose the execution route¶

direct = PanoramicDenseStereo(route="sphere-native")
result = direct(erp_a, erp_b, pose=known_pose)

Route	Geometry	Dense method	Return contract
`rectified-views`	paired rolled `GnomonicSpec` views	interchangeable classical matcher	`PanoramicStereoResult`
`sphere-native`	direct ERP epipolar curves	inverse-range plane sweep	`PanoramicStereoResult`

The routes are alternatives, not silently mixed estimates. The current sphere-native estimator has no source-validity argument, so this facade rejects partial source support on that route instead of treating invalid source pixels as evidence. The rectified-view route propagates validity for both panoramas through every projected pair.

Why the views already match¶

After pose resolution, the reference virtual camera’s local x-axis is aligned with the physical A-to-B baseline. Its source counterpart is the same axis rotated into camera B. Corresponding epipolar great circles therefore have the same raster row in both faces. Pair identity is geometric and deterministic; appearance does not decide which planned reference face belongs to which planned source face.

SphericalStereoViewPairer is the separate many-to-many utility for views that were sampled independently. Its score exposes angular overlap, sparse-inlier coverage, epipolar consistency and geometric support instead of hiding them in one opaque association.

Outside the application, arbitrary rolled views use the same object workflow:

views = panorama.project_views(specs)

for view in panorama.iter_views(specs):
    consume_then_release(view)

project_views() returns a GnomonicFaceSet; iter_views() projects only the next spec and is the appropriate route when view tensors must not coexist in memory.

Streaming and fusion¶

The projected route materializes one RGB pair and one cost volume at a time. Before allocating that volume, SelectStereoViewPairs records and gates joint support, texture, pose-inlier coverage, rectified-row residual and disparity interval viability. Accepted and rejected decisions remain in result.view_selection. The cost volume is discarded after disparity selection. Pair-local disparity is converted to radial range before backprojection; disparities from different pinhole lattices are never averaged. ERP fusion weights accepted range by matcher confidence and a face-edge taper, and reports weighted disagreement and contribution count.

The taper is only a geometric range-fusion weight. It is divided back out of the reported matching_confidence, so a correct observation does not appear less certain merely because it lies near a projected face boundary. consistency_confidence is exp(-disagreement / radial_range), and confidence is exactly the product of those two explicit components. None is a calibrated probability.

Finite pinhole views cannot cover the two epipolar singularities on the baseline axis. result.view_pairs and stage diagnostics expose the chosen FOV, pair count and omitted cone rather than claiming complete spherical coverage.

Cover the full sphere with multiple baselines¶

A cubemap fixes sampling coverage, but it cannot make the two epipoles of one camera pair observable. FullSphereDenseStereo therefore combines six overlapping cube-oriented tiles with two or more non-collinear registered source panoramas:

from panorai.applications import FullSphereDenseStereo
from panorai.stereo.panoramic import (
    PanoramicStereoPose,
    RegisteredPanoramaSet,
    StereoSourceView,
)

registered = RegisteredPanoramaSet(
    reference_id="A",
    reference=erp_a,
    sources=(
        StereoSourceView(source_id="B", panorama=erp_b, pose=pose_b_from_a),
        StereoSourceView(source_id="C", panorama=erp_c, pose=pose_c_from_a),
    ),
)

result = FullSphereDenseStereo()(registered)
range_m = result.radial_range

To reproduce the exact supplied-pose profile used by the promotion protocol, construct it by name instead of copying parameters:

stereo = FullSphereDenseStereo.reviewed_profile(filter_backend="auto")
result = stereo(registered)

assert stereo.describe()["pose_policy"] == "provided-only"
assert stereo.describe()["profile_id"].endswith("reviewed-2026-10-17")

This is a frozen candidate profile, not a Stable declaration. It covers only metric poses already present in RegisteredPanoramaSet. Automatic pose, scale recovery, multiscale AD-Census, tMGM, custom configurations and a probabilistic interpretation of confidence are outside that profile.

Registration belongs to the set: each source transform is interpreted only as T_source_from_A for this explicit reference. A source panorama is never described as independently registered.

For every face pixel and radial hypothesis, PanorAi forms the 3-D point in the reference frame, transforms it into each source camera, samples that source ERP, and aggregates only sources with sufficient parallax. Faces overlap by using a FOV greater than 90 degrees and are fused back into the reference ERP. Only one face and one hypothesis-sized source stack are retained at a time. The default per-hypothesis source aggregation is the median, which rejects one occluded or appearance-inconsistent source when at least three sources are available. weighted-mean and best remain explicit diagnostic alternatives.

Choose the cost pipeline explicitly¶

Full-sphere matching exposes three separate stages instead of hiding them behind one method name:

from panorai.applications import FullSphereDenseStereo
from panorai.stereo.panoramic import FullSphereStereoOptions

stereo = FullSphereDenseStereo(
    FullSphereStereoOptions(
        matching_cost="ad-census",  # raw unary evidence
        local_aggregation="box",    # one validity-normalized window
        regularizer="sgm4",         # four-path spatial regularization
    )
)
result = stereo(registered)

matching_cost is intensity-gradient, ad-census, or multiscale-ad-census; local_aggregation is none or box; and regularizer is none, sgm4, or tmgm. The stages are composed exactly once in that order. The selected configuration and its parameters are preserved in result.diagnostics and in the application’s saved JSON configuration.

The reviewed default is ad-census + box + sgm4, with angular ERP and face taps. Explicit matching_cost="intensity-gradient", erp_feature_tap_deg=None, and face_cost_tap_deg=None retain the historical profile. This selects the strongest frozen cross-domain candidate; it does not make the method Stable. The application remains Experimental.

Use the two-scale cost without changing the downstream aggregation:

multiscale = FullSphereDenseStereo(
    FullSphereStereoOptions(matching_cost="multiscale-ad-census")
)

This averages AD-Census evidence at the configured angular tap spacing and at twice that spacing. A frozen twelve-pair, view-disjoint audit excluded sky/no-return and non-observable directions before scoring. Relative to the single-scale route, fixed δ<1.25 changed from 0.345 to 0.349 on Stanford and from 0.440 to 0.445 on Matterport; conditional gross error fell by 0.008 and 0.009, while coverage fell by 0.003 and 0.005. Median paired runtime was 1.11x. The complete run reproduced bit for bit. The option is useful when the small accuracy/outlier gain justifies the additional work; it is not a calibrated confidence or a Stable-method claim.

A subsequent outcome-blind public multipair audit froze five Stanford 2D-3D pairs and six Matterport3D pairs before dense inference. It used the unchanged default above, a fixed evaluation population, explicit no-return exclusion, same-surface visibility, and solid-angle weighting:

Dataset	Pairs	Micro coverage	Micro δ<1.25	Median pair median error	Worst pair coverage / δ<1.25
Stanford 2D-3D	5	0.568	0.301	0.336 m	0.332 / 0.160
Matterport3D	6	0.536	0.339	0.192 m	0.388 / 0.195

Invalid predictions count as failures in δ<1.25; coverage is not removed from the denominator. A complete replay reproduced every metric and all eleven NPZ payloads bit for bit. Reproducibility is therefore established for this source environment, but accuracy and worst-case coverage still reject Stable promotion.

The same frozen audit tested whether the current confidence margin at least ranks predictions usefully. Retaining a predetermined fraction of each pair’s valid predictions gave:

Dataset	Retained predictions	Fixed-population coverage	Conditional δ<1.25
Stanford 2D-3D	25%	0.142	0.827
Stanford 2D-3D	50%	0.284	0.723
Stanford 2D-3D	100%	0.568	0.529
Matterport3D	25%	0.134	0.835
Matterport3D	50%	0.268	0.764
Matterport3D	100%	0.536	0.632

All eleven pairs were monotonic: the highest-confidence subset was more accurate than larger retained subsets. This supports confidence as a ranking diagnostic, but not as a calibrated probability or a portable threshold. The 25% subset covers only about 14% of the fixed observable population. A Stable abstention policy still requires a threshold learned without held-out outcomes and then tested unchanged on separate pairs.

That absolute-threshold test was then run at a frozen operating target: conditional δ<1.25 of at least 0.75 and fixed-population coverage of at least 0.10. For each fold, the most permissive absolute score satisfying both constraints was selected from calibration pairs only and applied unchanged to the target:

Calibration route	Joint passes	Main failure
same-domain leave-one-pair-out	4 / 11	seven target pairs fell below 0.75 accuracy or 0.10 coverage
all-other-pairs leave-one-pair-out	4 / 11	the pooled threshold did not preserve target accuracy
Stanford → Matterport	pass	0.837 conditional accuracy at 0.226 fixed coverage
Matterport → Stanford	fail	0.655 conditional accuracy at 0.377 fixed coverage

The two source replays produced identical thresholds and evaluations. The failure is therefore not nondeterminism: the numeric confidence scale is not portable between all pairs or both domain directions. Use it to rank evidence within a result, but do not expose a universal threshold or call it calibrated. Promotion requires a revised score or an independently frozen calibration model tested prospectively on unseen scenes.

The taper correction preserved radial range, validity, disagreement, support, observability and all source-count fields bit for bit on all eleven pairs. It improved within-pair ranking, but did not improve absolute-threshold transfer. Geometric blending contamination was therefore a real defect, not the main calibration failure.

A separate prospective multiview audit then froze three two-source references per public domain using only metric baseline and registered-cloud overlap. No dense output entered selection, sky/no-return directions were excluded, and the single- and two-source routes were scored on the same observable-union population:

Dataset / route	Micro coverage	Micro δ<1.25	Conditional gross error (>2 inverse-range labels)
Stanford / primary source	0.500	0.208	0.499
Stanford / two sources	0.528	0.233	0.482
Stanford / strict two-source consensus	0.041	0.028	0.213
Matterport / primary source	0.637	0.362	0.338
Matterport / two sources	0.655	0.439	0.253
Matterport / strict two-source consensus	0.100	0.082	0.102

Two sources improved fixed-population δ<1.25 in five of six triplets and reduced the conditional gross-outlier rate in both domain aggregates. This is real evidence for multibaseline aggregation, but the remaining Stanford error and low absolute accuracy still reject Stable promotion. Requiring both source views to choose a nearby lattice label produces a cleaner subset but collapses coverage to 4–10%; it is therefore an audit/abstention mode, not the default visibility solution. Both complete executions reproduced the deterministic metrics and all eighteen array payload hashes exactly.

Replacing the two-source lower median with parallax-weighted mean was tested unchanged on the same frozen triplets as a subsequent ablation. It reduced observable-union micro δ<1.25 from 0.233 to 0.197 on Stanford and from 0.439 to 0.388 on Matterport; Stanford coverage also fell from 0.528 to 0.482. Requiring every visible source to contribute therefore does not solve occlusion. The default is unchanged: a future policy must reject an occluded/incompatible source while still penalizing unsupported minima from a single ambiguous source.

A second parameter-free ablation multiplied the existing confidence by (1 + source_consensus_count / source_count) / 2. It changed neither range nor validity. On the same six triplets, however, median top-quartile conditional δ<1.25 decreased from 0.744 to 0.740 on Stanford and from 0.835 to 0.825 on Matterport. Absolute-threshold transfer was unchanged: only one of six all-other-pair folds and one of two cross-domain directions passed the frozen accuracy-and-coverage gate. Source-label agreement is therefore useful as a separate diagnostic, but this simple multiplication is rejected as a confidence improvement and is not part of the runtime score.

Selecting one fixed source group from complete per-source cost curves was also tested without using evaluation range. The most reliable source anchored the group, independently preferred labels had to agree within the existing lattice tolerance, and a second streaming pass formed a parallax-weighted mean. This prevented source switching between labels and reduced Stanford’s conditional gross-error fraction from 0.482 to 0.460. It nevertheless reduced Stanford coverage/δ<1.25 from 0.528/0.233 to 0.496/0.222, and worsened Matterport from 0.655/0.439/0.253 to 0.638/0.417/0.267 for coverage/δ<1.25/gross-error. The policy was therefore rejected and was not added to the runtime API. Visibility cannot be inferred reliably from unary curve sharpness alone.

A prospective three-source protocol then selected six new four-view sets from pose and registered-cloud overlap before inference. The unchanged lower median became a true median when three costs were available:

All values below exclude sky and no-return directions before scoring. The fixed denominator is the prediction-independent intersection of positive, finite dataset range, geometric sampling support and the stated observability population. A missing or rejected prediction inside that denominator still counts against coverage and fixed-population accuracy.

Dataset / sources	Coverage	Fixed-population δ<1.25	Conditional gross >2-label error
Stanford / one	0.422	0.150	0.533
Stanford / two	0.468	0.195	0.490
Stanford / three	0.430	0.178	0.464
Matterport / one	0.704	0.413	0.341
Matterport / two	0.728	0.515	0.238
Matterport / three	0.714	0.514	0.224

The third source suppresses gross outliers in both domains. It does not yet improve the fixed-denominator result because accepted coverage falls. On the population visible in all three sources, Matterport δ<1.25 increases from 0.537 to 0.542, while Stanford decreases from 0.225 to 0.213. Requiring three valid source costs per hypothesis produces almost the same result as the ordinary three-source route, so missing source support is not the main explanation.

A frozen gate-decomposition then disabled one existing acceptance condition at a time without choosing replacement thresholds from outcomes. The table shows the three-source result minus the two-source result on the same observable-union population:

Domain / policy	Δ coverage	Δ fixed δ<1.25	Δ gross fixed fraction
Stanford / baseline	-0.0385	-0.0172	-0.0299
Stanford / no cost gate	-0.0385	-0.0172	-0.0299
Stanford / no confidence gate	-0.0274	-0.0160	-0.0227
Stanford / no texture gate	-0.0502	-0.0387	-0.0253
Matterport / baseline	-0.0144	-0.0017	-0.0131
Matterport / no cost gate	-0.0144	-0.0017	-0.0131
Matterport / no confidence gate	-0.0054	+0.0002	-0.0079
Matterport / no texture gate	-0.0387	-0.0122	-0.0269

The absolute matching-cost gate is inactive on all 60 evaluated conditions: setting it from 0.65 to 1.0 changes no output. Confidence explains part of the source-count coverage loss, especially on Matterport, but not Stanford. Texture controls much of the absolute coverage; it does not cause the two-to-three-source regression. The next correction must therefore instrument rejection reasons after cost fusion—especially confidence and boundary-label rejection—before defining a source-count-aware score prospectively. Relaxing the current cost threshold cannot solve the observed problem.

The estimator now exposes the cumulative face-pixel counts that produced this decision in result.diagnostics["face_acceptance_cumulative_counts"]. These counts follow the declared face_acceptance_order; overlapping faces are counted separately, so they diagnose the algorithm but are not ERP accuracy metrics. Replaying the unchanged VAL-112 protocol preserved all 24 numerical array hashes bit for bit and produced:

Domain / cumulative stage	Two sources	Three sources	Difference
Stanford / interior hypothesis	266996	253836	-13160
Stanford / after texture	153346	144737	-8609
Stanford / after confidence = accepted	135211	123309	-11902
Matterport / interior hypothesis	284052	281350	-2702
Matterport / after texture	224525	224294	-231
Matterport / after confidence = accepted	212276	208775	-3501

Stanford’s regression is primarily associated with the median choosing an endpoint of the finite inverse-range lattice; such endpoints cannot support the current parabolic sub-hypothesis refinement and are conservatively rejected. Matterport’s remaining regression is primarily the normalized best-versus-second-best confidence gate. Matching cost, minimum source count, range consensus and the final parallax predicate remove no additional face samples after confidence in this frozen configuration. This does not justify accepting endpoint labels or dropping confidence: both alternatives need independent correctness oracles and prospective public validation.

That validation was then run with the exact discrete endpoint label accepted without parabolic refinement. The candidate reduced the two-to-three-source coverage gap, but failed every frozen promotion criterion:

Domain / three-source policy	Coverage	Fixed δ<1.25	Conditional gross error
Stanford / reject endpoint	0.4296	0.1782	0.4639
Stanford / accept discrete endpoint	0.4810	0.1772	0.5135
Matterport / reject endpoint	0.7139	0.5136	0.2242
Matterport / accept discrete endpoint	0.7495	0.5136	0.2566

Coverage increased, but fixed-population accuracy decreased slightly in both domains and conditional gross error increased by 0.0496 on Stanford and 0.0323 on Matterport, above the frozen 0.02 limit. Two full executions reproduced the evidence and all 24 payload hashes exactly. The candidate option was removed: a boundary winner must remain rejected until the estimator can distinguish a true endpoint from a censored out-of-range or ambiguous match. The principled next routes are range-interval expansion with a second search, or an explicit bound/censored output contract—not silently treating the bound as a point range.

Confidence was then separated prospectively from hard validity on twelve new public pairs: six Stanford pairs balanced across areas 2, 3 and 4, and six Matterport pairs from distinct spatial groups. The manifest was frozen before inference, excluded every pair and view used by the earlier protocols, and used only pose, registered-cloud overlap and asset availability. Evaluation again removed sky/no-return with positive finite dataset range before applying prediction validity.

Domain / confidence policy	Coverage	Fixed δ<1.25	Conditional gross error
Stanford / threshold	0.5045	0.3032	0.3666
Stanford / report only	0.5630	0.3140	0.4105
Matterport / threshold	0.6412	0.4881	0.2165
Matterport / report only	0.6754	0.4995	0.2380

Report-only confidence increased coverage and fixed-population accuracy in both domains, and confidence selectivity remained monotonic. It nevertheless increased Stanford conditional gross error by 0.04386, exceeding the frozen 0.04 tolerance; Matterport increased by 0.02150. Two complete executions produced identical evidence and all 24 result payload hashes. The candidate failed as specified, was removed, and the threshold default is unchanged. This isolates the next requirement: improve the score or its independently validated acceptance rule rather than making confidence advisory everywhere.

A separate cost-default protocol then froze twelve further view-disjoint pairs before inference and held box aggregation, SGM4, geometry, range lattice and all acceptance predicates constant. It compared only the unary matching cost:

Domain / unary cost	Coverage	Fixed δ<1.25	Conditional gross error
Stanford / intensity-gradient	0.5888	0.2849	0.4963
Stanford / AD-Census	0.6164	0.3395	0.4379
Matterport / intensity-gradient	0.6312	0.4196	0.2795
Matterport / AD-Census	0.6353	0.4228	0.2803

AD-Census passed every frozen scientific criterion: fixed-population accuracy improved in both domains, coverage did not regress, and conditional gross error stayed within the 0.02 tolerance. Median runtime was 0.923× the existing cost in the first run and 0.904× in the replay, below the 1.25× limit. The deterministic evidence and all 24 result payloads reproduced exactly. A fresh review reproduced all payloads, populations, metrics, and aggregates. The raw review JSON additionally contains the later angular-neighbourhood diagnostic; removing only that field reconstructs the original evidence byte for byte. AD-Census is consequently the default for new option objects. Old saved configurations without the field still deserialize to the historical cost.

A resolution/lattice diagnostic then reused the already observed VAL-098 development pairs. It is not held-out promotion evidence. A controlled 512x1024 ERP factorial kept AD-Census, box aggregation, SGM4, poses, populations and every acceptance gate fixed while varying face size and the inverse-range lattice:

Controlled cell	Stanford coverage / fixed δ<1.25	Matterport coverage / fixed δ<1.25
face 128 / 48 hypotheses	0.4361 / 0.2525	0.4197 / 0.2775
face 128 / 96 hypotheses	0.4446 / 0.2837	0.4121 / 0.2785
face 256 / 48 hypotheses	0.4424 / 0.2176	0.4210 / 0.2440
face 256 / 96 hypotheses	0.4543 / 0.2568	0.4186 / 0.2559

Doubling the range lattice improves fixed accuracy by 0.0312 on Stanford but only 0.0010 on Matterport. Doubling face resolution with the unchanged pixel-window parameters lowers fixed accuracy by about 0.034 in both domains. The result reproduces exactly for the best controlled cell and rules out a simple “use more pixels” remedy. It also exposes a contract issue: the same pixel-valued window_size currently controls an ERP feature neighbourhood and projected-face cost aggregation, so its angular support changes when ERP or face resolution changes. Angularly invariant feature, Census and aggregation support must be evaluated before a higher-resolution configuration can be a Stable default.

The follow-up implementation separates those roles without changing legacy execution when the new options are omitted:

from panorai.stereo.panoramic import FullSphereStereoOptions

angular = FullSphereStereoOptions(
    matching_cost="ad-census",
    erp_feature_tap_deg=0.703125,
    face_cost_tap_deg=1.0667900073,
)

erp_feature_tap_deg is passed to the tangent-plane spherical feature filter. face_cost_tap_deg is converted from actual adjacent centre-ray spacing to a deterministic integer stride and is shared by Census and valid-normalized local aggregation. The number of Census bits and aggregation taps stays constant; only their represented angle is preserved. None retains the legacy one-pixel tap exactly.

On the same eleven development pairs at 512x1024, face 256 and 96 range hypotheses, preserving both angular supports improved every aggregate domain:

Neighbourhood	Stanford coverage / fixed δ<1.25	Matterport coverage / fixed δ<1.25
legacy pixel taps	0.4543 / 0.2568	0.4186 / 0.2559
angular ERP only	0.5973 / 0.3273	0.5506 / 0.3301
angular face only	0.4490 / 0.2883	0.4150 / 0.2789
angular ERP + face	0.5910 / 0.3694	0.5456 / 0.3647

The complete route gains about 0.11 fixed δ in both domains and reproduces all eleven candidate payload hashes exactly. A prospectively frozen follow-up then selected twelve new view-disjoint pairs without inspecting RGB matches, dense output, or range-reference outcomes. The fixed population excluded sky/no-return through the dataset’s positive finite range return, then required geometric support and independent same-surface observability. Prediction validity never changed this denominator.

Held-out domain / neighbourhood	Coverage	Fixed δ<1.25	Conditional error >2x
Stanford / legacy pixels	0.4351	0.2281	0.2810
Stanford / angular ERP + face	0.5723	0.3321	0.2391
Matterport / legacy pixels	0.5301	0.4130	0.1095
Matterport / angular ERP + face	0.6420	0.5130	0.0940

The angular route passed every frozen scientific gate, improved all three reported measures in both domains, and took 1.226× the median runtime in the first execution and 1.212× in the replay, below the 1.35× ceiling. The complete evidence file, all 24 output payloads, and all 24 population masks reproduced exactly. A separate reviewer then reproduced the same evidence SHA and every payload/population hash. The angular values are now defaults for new option objects, while explicit None and old saved configurations preserve the legacy one-pixel behavior. This does not promote panoramic dense stereo to Stable: pose-estimation validation, transferable uncertainty/abstention evidence, and installed-wheel streaming evidence remain open.

The three masks answer different questions:

sampling_support: did at least one cube tile cover this direction?
observable: did some configured radial hypothesis have enough registered baselines with valid source support and parallax? This geometric mask is independent of the appearance cost and of the range selected by that cost.
validity: did the matching cost, texture, confidence, and range interval also pass?

observable_source_count is the largest number of geometrically usable source views found at one configured radial hypothesis. source_count instead counts the cost-valid sources at the selected range. Keeping both prevents Census or another descriptor’s border support from being misreported as missing stereo geometry.

source_consensus_count answers another question: how many source views independently preferred an inverse-range label within the configured lattice tolerance of the fused label? It is diagnostic by default. Set min_range_consensus_sources=2 only when a conservative two-source subset is more valuable than dense coverage; the rejected pixels remain invalid rather than being filled from another prior.

Thus full-sphere sampling does not imply fabricated full-sphere depth. Occlusion, missing source support, textureless regions, and inconsistent appearance can still leave valid directions empty. With only one source, directions around its baseline axis remain explicitly unobservable.

Inspect or replace stages¶

pipeline = stereo.as_pipeline()
print(tuple(pipeline.named_steps))

The projected route exposes:

ResolveStereoPose;
ResolveStereoScale;
PlanRectifiedStereoViews;
ProjectStereoViewPairs;
SelectStereoViewPairs;
DenseStereoPerPair;
BackProjectStereoRange;
FuseStereoRange;
ValidatePanoramicStereo.

Use the low-level APIs when developing a new matcher or pairing policy. Use the application façade when the desired output is a complete reference-ERP range product.