Validate and consume radial range

Level: Evaluate · Applies to: stereo and monocular outputs

Before creating points or surfaces, record and check:

  1. Quantity: PanorAi expects radial range rho, where X = rho * bearing; pinhole Z-depth is a different value.

  2. Units and scale: identify the baseline, model calibration or other source of metric scale.

  3. Validity: retain an explicit mask; zero and black RGB are not invalidity signals.

  4. Support: distinguish unobserved directions from rejected predictions.

  5. Consistency: inspect reprojection, left/right or cross-view agreement, confidence and seam behaviour.

  6. Provenance: retain model/checkpoint identity, pose, sampling, conversion and fusion configuration.

Only accepted pixels should become X = range * ray. Compare against scanner range or another independent reference when making accuracy claims.

Quantitative checks

When reference range is available, report error over the intersection of reference validity, prediction validity and geometric support. Include robust statistics and break them down by latitude, seam distance, range band and surface discontinuity. Coverage must accompany error: rejecting difficult pixels can improve RMSE without improving the product.

For stereo, also report bidirectional agreement and reprojection residual. For monocular methods, report scale alignment separately from shape error unless the checkpoint is explicitly metric. Never fit scale on the evaluation pixels and then describe the result as an unadjusted metric prediction.

3-D sanity check

Compute rays from the same canonical ERP convention and form points = range[..., None] * rays. Verify a small set of known directions and distances before exporting a cloud. A transposed ERP, axial/radial mix-up or wrong camera frame can produce a visually plausible but geometrically wrong surface.

Continue to multiview reconstruction or the 3-D reconstruction guide.