Residual-consensus driven linear matching

We propose a residual-consensus driven linear matching algorithm for simultaneous geometric parameter and point correspondence estimation. Using the linearization technique, we quantize geometric transformation into discrete levels with regard to each correspondence matrix. We identify the uncontaminated models by evaluating the statistical coherent of residual ordering and Maximum mean discrepancy (MMD) test. The uncontaminated models provide the optimal quantization coefficients for each assignment matrix of our linear programming model. Experiment results on a variety of images demonstrate our method effectiveness and robustness.