FSASAC: Random Sample Consensus Based on Data Filter and Simulated Annealing

Verification with RANSAC has become a crucial step for local feature-based matching applications in machine vision. Details of its implementation are directly relevant to time consumption and the quality of the estimated results. A novel method of robust model estimation for matched images with a large viewpoint change, namely Random Sample Consensus Based on Data Filter and Simulated Annealing, FSASAC in short, is introduced. FSASAC begins by filtering out most outliers according to their distributions, and then adjusts the sample probability of matched pairs by Simulated Annealing in the process of iteration. The process of sampling ceased until the current best hypothesis met with the termination conditions. Finally, FSASAC is presented experimentally, both on synthesized data and image pairs. Choose some modified methods as the comparative objectives. The results indicate that FSASAC can estimate the optimal model from image pairs with lower number of iterations and run-time under the condition of large viewpoint change, even when the outlier ratio is higher than 90%.