[Feature]: Explore RANSAC variants for two view pose estimation

🚀 Feature Description

Add support for advanced RANSAC variants in the kornia-3d pose estimation module, specifically:

1. LO-RANSAC (Locally Optimized RANSAC): Performs local optimization on promising hypotheses to improve model quality before final selection.

2. GC-RANSAC (Graph-Cut RANSAC): Uses graph-cut optimization for spatial consistency in inlier selection, improving robustness in structured scenes.

3. PROSAC (Progressive Sample Consensus): Leverages match quality scores for more efficient sampling, reducing iterations needed for convergence.

These variants would enhance the existing ransac_fundamental and ransac_homography functions in kornia-3d/src/pose/twoview.rs, as well as the PnP RANSAC in kornia-3d/src/pnp/ransac.rs.

📂 Feature Category

Geometry

💡 Motivation

The current RANSAC implementation in kornia-3d uses a basic vanilla RANSAC approach. While functional, this has limitations:

1. Efficiency: Vanilla RANSAC may require many iterations to find a good model, especially with high outlier ratios.

2. Model Quality: Without local optimization, the final model may not be as refined as it could be.

3. Spatial Consistency: Vanilla RANSAC treats points independently, missing opportunities to leverage spatial structure.

As discussed by @ducha-aiki in PR #688:

"yes - LO-RANSAC, GC-RANSAC. Overall I'd mimic https://github.com/PoseLib/PoseLib"

💭 Proposed Solution

1. Extend RansacParams Configuration

Add configuration options to support different RANSAC strategies:

/// RANSAC method variants
#[derive(Clone, Copy, Debug, Default)]
pub enum RansacMethod {
    /// Standard RANSAC (current implementation)
    #[default]
    Standard,
    /// Locally Optimized RANSAC - refines promising hypotheses
    LoRansac {
        /// Number of local optimization iterations
        lo_iterations: usize,
        /// Inlier threshold multiplier for LO step
        lo_threshold_mult: f64,
    },
    /// Graph-Cut RANSAC - spatial consistency via graph cuts
    GcRansac {
        /// Neighborhood radius for graph construction
        neighborhood_radius: f64,
        /// Spatial coherence weight
        spatial_weight: f64,
    },
    /// Progressive RANSAC - quality-guided sampling
    Prosac {
        /// Maximum PROSAC-specific iterations before falling back to uniform sampling
        max_prosac_iterations: usize,
    },
}

#[derive(Clone, Debug)]
pub struct RansacParams {
    /// Maximum number of RANSAC iterations
    pub max_iterations: usize,
    /// Minimum number of iterations before early termination
    pub min_iterations: usize,
    /// Inlier threshold (pixel error)
    pub threshold: f64,
    /// Minimum number of inliers required for acceptance
    pub min_inliers: usize,
    /// Target success probability for adaptive iteration count
    pub confidence: f64,
    /// RANSAC method variant
    pub method: RansacMethod,
    /// Optional RNG seed for deterministic runs
    pub random_seed: Option<u64>,
}

📚 Library Reference

PoseLib (C++): https://github.com/PoseLib/PoseLib

🔄 Alternatives Considered

No response

🎯 Use Cases

1. Visual SLAM Initialization: Two-view initialization with robust pose recovery from potentially noisy feature matches

2. Structure from Motion: Essential/fundamental matrix estimation in SfM pipelines with varying outlier ratios

📝 Additional Context

No response

🤝 Contribution Intent

• I plan to submit a PR to implement this feature
• I'm requesting this feature but not planning to implement it