Compare non-learned stereo pipelines

Level: Compose/evaluate · Input: an already-rectified pinhole pair · Output: disparity, validity, confidence and optional radial range

These Experimental baselines make every major classical stereo step inspectable. Their input is one already-rectified pinhole pair; they do not interpret an ERP. Use them directly when studying a matcher. Use PanoramicDenseStereo to resolve pose, plan and project rectified view pairs, run the selected matcher and fuse radial range back onto the sphere.

from panorai.applications import StereoSGBMDepth
from panorai.stereo.classical import (
    RectifiedStereoGeometry,
    RectifiedStereoPair,
    StereoSGBMConfig,
)

stereo = StereoSGBMDepth(
    StereoSGBMConfig(min_disparity=0, max_disparity=128)
)
result = stereo(
    rectified_a,
    rectified_b,
    reference_support=support_a,
    source_support=support_b,
    geometry=RectifiedStereoGeometry(
        focal_length_x=fx_px,
        focal_length_y=fy_px,
        principal_point_x=cx_px,
        principal_point_y=cy_px,
        baseline=baseline_m,
    ),
)

disparity_px = result.disparity
radial_range_m = result.radial_range
valid = result.validity
rectified_pair = RectifiedStereoPair(
    rectified_a, rectified_b, support_a, support_b
)

disparity is $x_A-x_B$ in pixels. Geometry first gives optical-axis depth $z=f_xB/d$ and then returns radial range from the reference camera centre. Invalid numeric samples are NaN; validity, occlusion and mismatch are separate masks. Zero intensity and zero-valued evidence remain valid data.

Six baselines

Application	Inspectable composition	Implementation boundary
StereoBMDepth	block matching → validation	OpenCV StereoBM
ADCensusStereoDepth	AD-Census → cross aggregation → SGM → selection → validation	PanorAi reference operations
StereoSGBMDepth	block evidence + full SGM → validation	OpenCV StereoSGBM HH
CensusMGMStereoDepth	Census → box aggregation → MGM or tMGM → selection → validation	PanorAi reference operations
ELASStereoDepth	Census → aggregation → triangulated support prior → SGM → selection → validation	ELAS-style PanorAi reference; not binary-equivalent to libELAS
PatchMatchStereoDepth	slanted-plane PatchMatch → validation	deterministic PanorAi reference

All six wrappers expose the same call and return type. Select a bounded disparity interval appropriate to the rectified baseline; the defaults are not a scene-specific calibration.

run = stereo.as_pipeline().run(
    rectified_pair,
    retain=("matching_and_optimization",),
)
print(run.describe())

For bottom-up experiments, compose the same public operations directly:

from panorai.pipeline import Pipeline
from panorai.stereo.classical import (
    ADCensusCost,
    BuildStereoCost,
    CrossBasedCostAggregation,
    RefineStereoDisparity,
    SelectStereoDisparity,
    SemiGlobalMatching,
    TransformStereoCost,
)

pipeline = Pipeline(
    BuildStereoCost(ADCensusCost(), 0, 128),
    TransformStereoCost(CrossBasedCostAggregation(), "aggregation"),
    TransformStereoCost(SemiGlobalMatching(paths=8), "optimization"),
    SelectStereoDisparity(subpixel=True),
    RefineStereoDisparity(method="my-ad-census"),
)
result = pipeline(rectified_pair)

The internal cost volume is $H\times W\times D$ and can be large. MGM/tMGM, the ELAS-style reference and PatchMatch are correctness-oriented reference implementations, not current performance claims. Benchmark after fixing image size, disparity range, support, warm-up, repetitions and hardware.

What is independently checked

The focused regression uses a textured rectified pair with a closed-form translation and known disparity. It verifies all six compositions, explicit support, radial-range conversion, immutable results and safe application configuration. This is an analytic correctness gate, not evidence that any method is superior on real scenes. Wide-baseline exposure change, repeated industrial texture, occlusion and rectification error remain prospective evaluation requirements.