Adaptive Robust Homography Estimation via Hybrid RANSAC and Iteratively Reweighted Least Squares

Abstract

Homography estimation under heavy illumination variation and viewpoint changes remains a challenging problem because of high outlier ratios and noise with non-Gaussian statistics. This paper introduces the Adaptive Hybrid Iteratively Reweighted Least Squares (AH-IRLS) algorithm, which leverages both robust initialization via RANSAC and iterative refinement through the IRLS procedure.Two adaptive mechanisms are embedded in the algorithm.First, the inlier threshold is determined automatically using the Median Absolute Deviation (MAD), eliminating the need for any manual parameter setting. Second, at every iteration, the robust loss function is selected adaptively based on the skewness and kurtosis of the current residual distribution. The core idea is not to invent new estimation primitives but to combine established robust statistical tools intelligently within a single homography estimation pipeline. Numerical stability is ensured through Hartley normalization and weightedsingular-value decomposition (SVD).Experiments on the HPatches benchmark show consistent improvements over classical RANSAC. Advanced sampling-based baselines (PROSAC, MAGSAC++) are also considered for additional validation. On the v grace and i toy sequences, AH-IRLS reduces the root-mean-square error (RMSE) by 56.8\,\% and 52.8%, respectively, while achieving a mean inlier rate of 94.2%. These results confirm that combining adaptive statistical modeling with robust geometric estimation is highly effective for challenging real-world conditions.