Model fitting is a fundamental component in\ncomputer vision for salient data selection, feature extraction\nand data parameterization. Conventional approaches such as\nthe RANSAC family show limitations when dealing with\ndata containing multiple models, high percentage of outliers\nor sample selection bias, commonly encountered in computer\nvision applications. In this paper, we present a novel model\nevaluation function based on Gaussian-weighted Jensenââ?¬â??\nShannon divergence, and integrate into a particle swarm optimization\n(PSO) framework using ring topology. We avoid\ntwo problems from which most regression algorithms suffer,\nnamely the requirements to specify inlier noise scale and the\nnumber of models. The novel evaluation method is generic\nand does not require any estimation of inlier noise. The continuous\nand meta-heuristic exploration facilitates estimation\nof each individual model while delivering the number of\nmodels automatically. Tests on datasets comprised of inlier\nnoise and a large percentage of outliers (more than 90 %\nof the data) demonstrate that the proposed framework can\nefficiently estimate multiple models without prior information.\nSuperior performance in terms of processing time and\nrobustness to inlier noise is also demonstrated with respect\nto state of the art methods.
Loading....