Toward Reliable Homography Estimation under Adverse Degradations: An Optimization-Driven Approach

Area of Science:

• Computer Vision
• Geometric Computer Vision
• Image Processing

Background:

• Homography estimation is crucial for geometric computer vision but struggles with adverse imaging conditions.
• Degradation and appearance discrepancies in scenarios like low-light, haze, and rain weaken feature correlations, leading to unstable correspondences.
• Existing hand-crafted and data-driven methods are prone to errors under these challenging conditions.

Purpose of the Study:

• To propose an optimization-driven framework for robust homography estimation.
• To address the limitations of current methods in adverse imaging conditions.
• To improve alignment accuracy and generalization across diverse degradations.

Main Methods:

• Developed a progressive energy minimization framework for robust alignment.
• Introduced a Robust Feature Initialization (RFI) strategy for static-plane-aware warm starts.
• Incorporated an Alignment Fidelity Module (AFM) for progressive deformation refinement.
• Utilized a Perception Regularization Module (PRM) for task-driven semantic constraints.
• Implemented a Degradation Constraint Learning (DCL) strategy with an auxiliary restoration decoder.

Main Results:

• Achieved state-of-the-art alignment accuracy on a comprehensive harsh-environment homography benchmark.
• Demonstrated robust generalization across diverse degradations (low-light, haze, rain, etc.).
• Showcased consistent improvements on five downstream perception tasks.
• Maintained practical inference efficiency with an adaptive stopping criterion.

Conclusions:

• The proposed optimization-driven framework significantly enhances homography estimation reliability under adverse conditions.
• The complementary modules (RFI, AFM, PRM, DCL) effectively address feature correlation weaknesses.
• The method offers a robust and efficient solution for challenging geometric computer vision problems.