Skip to Main content Skip to Navigation

BetaSAC et OABSAC, deux nouveaux 'echantillonnages conditionnels pour RANSAC

Antoine Méler 1 
1 PRIMA - Perception, recognition and integration for observation of activity
Inria Grenoble - Rhône-Alpes, UJF - Université Joseph Fourier - Grenoble 1, INPG - Institut National Polytechnique de Grenoble , CNRS - Centre National de la Recherche Scientifique : UMR5217
Abstract : RANSAC algorithm (Random Sample Consensus) is the most common approach for the problem of robust parameters estimation of a model in computer vision. This is mainly its ability to handle data containing more errors than potentially useful information that made its success in this area where sensors provide a very rich but noisy information. Since its creation thirty years ago, many modifications have been proposed to improve its speed, accuracy and robustness. In this work, we propose to accelerate the resolution of a problem by using RANSAC with more information than traditional approaches. This information, extracted from the data itself or from complementary sources of all types, is used help generating more relevant RANSAC hypotheses. To do this, we propose to distinguish four degrees of quality of a hypothesis : inlier, consistent, coherent or suitable sample. Then we show how an inlier sample is far from being relevant in the general case. Therefore, we strive to design a novel algorithm which, unlike previous methods, focuses on the generation of suitable samples rather than inlier ones. We begin by proposing a probabilistic model unifying all the RANSAC reordered sampling methods. These methods provide a guidance of the random data selection without impairing the search. Then, we propose our own scheduling algorithm, BetaSAC, based on conditional partial sorting. We show that the conditionality of the sort can satisfy consistency constraints, leading to a generation of suitable samples in the first iterations of RANSAC, and thus a rapid resolution of the problem. The use of partial rather than exhaustive sorting ensures rapidity and randomization, essential to this type of methods. In a second step, we propose an optimal version of our method, called OABSAC (for Optimal and Adaptive BetaSAC), involving an offline learning phase. This learning is designed to measure the properties of the specific problem that we want to solve in order to determine automatically the optimal setting of our algorithm. This setting is the one that should lead to a reasonably accurate estimate of the model parameters in a shortest time (in seconds). The two proposed methods are very general solutions that integrate into RANSAC any additional useful information. We show the advantage of these methods to the problem of estimating homographies and epipolar geometry between two photos of the same scene. The speed gain compared to the classical RANSAC algorithm can reach a factor of hundred
Complete list of metadata

Cited literature [65 references]  Display  Hide  Download
Contributor : Dominique Vaufreydaz Connect in order to contact the contributor
Submitted on : Monday, January 27, 2014 - 10:36:37 AM
Last modification on : Friday, March 25, 2022 - 9:44:20 AM
Long-term archiving on: : Sunday, April 27, 2014 - 10:31:03 PM


  • HAL Id : tel-00936650, version 1



Antoine Méler. BetaSAC et OABSAC, deux nouveaux 'echantillonnages conditionnels pour RANSAC. Vision par ordinateur et reconnaissance de formes [cs.CV]. Université de Grenoble, 2013. Français. ⟨tel-00936650⟩



Record views


Files downloads