v0.1.2

Bug fixes

• LogPolarCorrelator: rotation estimation now actually works. Two correctness bugs in LogPolarCorrelator caused it to silently degrade to translation-only correlation in 0.1.1:
  • sampleMagnitude was treating the natural-layout FFT spectrum as fft-shifted, so the polar sampling rays emanated from the (-N/2, -N/2) Nyquist corner instead of from DC.
  • The dy → rotation conversion had the wrong sign, so even after the addressing was fixed the second-pass de-rotation rotated frame B further away from frame A rather than back toward it.
  • Validated against synthetic ground truth: rotation error ≤ 0.04° over the ±60° range, sub-pixel residual translation after de-rotation. Anyone using LogPolarCorrelator on 0.1.1 should upgrade.
• Removed a duplicated WGSL const declaration in the log-polar magnitude shader (silently shipped in 0.1.1).

New API

• PhaseCorrelator.correlateCartesian(imageA, imageB) and LogPolarCorrelator.correlateCartesian(imageA, imageB) accept pre-computed N×N Cartesian Float32Arrays directly, skipping the polar reprojection. Useful for testing the correlators in isolation against synthetic data, or for general image registration on already-Cartesian inputs.

Full Changelog: v0.1.1...v0.1.2

v0.1.1

Full Changelog: https://github.com/Arechii/webgpu-sonar/commits/v0.1.1