docs(motion): clarify stabiliser option and memory summaries

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
This commit is contained in:
Claude committed 2026-10-02 09:56:35 +00:00
1 parent 310379d2da
commit 17bdaeebf6
1 file changed
+9 -8
+9 -8
View File
@@ -23,9 +23,10 @@ public sealed record StabilizerOptions
/// <summary>Residual (px) under which a patch vector is an inlier of a similarity hypothesis.</summary>
public double InlierThreshold { get; init; } = 1.0;
/// <summary>Largest plausible frame-to-frame rotation (radians) and |ln scale| of a hypothesis.</summary>
/// <summary>Largest plausible frame-to-frame rotation (radians) of a hypothesis.</summary>
public double MaxRotation { get; init; } = 0.05;
/// <summary>Largest plausible frame-to-frame |ln scale| of a hypothesis.</summary>
public double MaxScaleChange { get; init; } = 0.05;
/// <summary>σ (frames) of the Gaussian-weighted local linear regression that smooths the camera path; 0 disables.</summary>
@@ -58,13 +59,13 @@ public sealed record StabilizationResult(
/// <summary>
/// Video stabilisation. Per frame: Kuglin–Hines phase correlation of nine patches on a 3×3 grid against the
/// previous frame (patches analysed in parallel), texture gating (noise-compensated gradient energy, Immerkær 1996 noise estimate; a
/// textureless frame yields no vectors and therefore no motion), then a robust similarity fit (exhaustive
/// two-point hypotheses as deterministic RANSAC, plausibility limits, weighted least squares on the
/// inliers). <see cref="Solve"/> accumulates the camera path, smooths its translation, rotation and log-scale
/// with Gaussian-weighted local linear regression (pans survive unchanged), and reports per-frame
/// corrections (smoothed ∘ raw⁻¹), shake RMS and the crop needed to hide borders. Memory holds two
/// luminance planes regardless of sequence length.
/// previous frame (patches analysed in parallel), texture gating (noise-compensated gradient energy,
/// Immerkær 1996 noise estimate; a textureless frame yields no vectors and therefore no motion), then a
/// robust similarity fit (exhaustive two-point hypotheses as deterministic RANSAC, plausibility limits,
/// weighted least squares on the inliers). <see cref="Solve"/> accumulates the camera path, smooths its
/// translation, rotation and log-scale with Gaussian-weighted local linear regression (pans survive
/// unchanged), and reports per-frame corrections (smoothed ∘ raw⁻¹), shake RMS and the crop needed to hide
/// borders. Memory is two luminance planes plus one small record per frame.
/// </summary>
public sealed class Stabilizer : IDisposable
{