PESAC, the generalized framework for RANSAC-based methods on SIMD computing platforms

This paper focuses on the computational optimization of RANSAC. We describe Parallel Efficient Sample Consensus (PESAC) framework that allows efficient utilization SIMD extensions and provides memory locality due to a special way storing input sequence correspondences generating batch samples per one main loop iteration. It is inspired by USAC has block structure capable implementing most modern RANSAC-based methods. enhance it with individual blocks sample model restrictors are aimed at rejection “bad” hypothesis before time-consuming computation verification blocks. also provide detailed description 2D homography estimation problem in PESAC benchmark running time MIDV-2020 dataset identity documents. Comparing naive implementation, we accelerated our 122 times for document classification task (with 6 % increase accuracy) 18 tracking 46 decrease failure rate) using both vector processing. version outperformed number implementations from OpenCV-4.6.0 runtime accuracy (3 faster, greater task, 2 33 lower rate if comparing USAC_MAGSAC).