Detect features directly on the sphere

Level: Extend · Strategy: Sphere-native detection with local projected description · Stability: Experimental

SphericalDoGSIFTPipeline builds its Gaussian and DoG scale spaces with spherical operations. Extrema are compared in tangent neighbourhoods, so the detector crosses the ERP seam without creating overlapping faces.

The current descriptor remains hybrid: PanorAi opens one small gnomonic patch at each accepted bearing and asks OpenCV SIFT to describe its centre. The patch does not decide where the keypoint exists.

Use this route when constant angular scale and seam continuity matter enough to justify an Experimental detector. Compare repeatability, high-latitude behaviour, runtime and pose quality against the Stable face-based route.

Composition from core operations

The detector is built from spherical Gaussian smoothing, scale changes, difference images and tangent-neighbourhood extrema. Pointwise subtraction and thresholding remain ordinary NumPy/Torch operations. Only steps that interpret spatial neighbourhoods need spherical implementations.

This separation matters: direct spherical detection avoids face-boundary duplicates, while projected description reuses a mature descriptor without pretending that SIFT’s planar orientation model was rewritten. The output is still the same feature-set contract used by the projected route, so matching and pose code do not depend on the detector topology.

Compare routes fairly

Hold the descriptor, matching policy and pose thresholds fixed. Report repeatability under longitude rotation, seam-crossing retention, latitude distribution, descriptors per steradian, runtime and final pose acceptance. More detected points alone is not evidence of a better spherical detector.

See composing spherical models for core composition and the features tutorial for the full comparison. Next: matching and diagnostics.