Automatic Threshold RanSaC Algorithms for Pose Estimation Tasks - Archive ouverte HAL Accéder directement au contenu
Article Dans Une Revue Communications in Computer and Information Science Année : 2023

Automatic Threshold RanSaC Algorithms for Pose Estimation Tasks

Résumé

When faced with data corrupted by both noise and outliers robust estimation algorithms like RanSaC are used, especially in the field of Multi-View Stereo (MVS) or Structure-from-Motion (SfM). To find the best model fitting the data, from numerous minimal samples it evaluates models and rates them according to their number of inliers. The classification as inlier depends on a user-set threshold that should be tailored to the noise level of the data. This requires the knowledge of this level, which is rarely available. The few existing adaptive threshold algorithms solve this problem by estimating the value of the threshold while computing the best model. However, it is hard to obtain groundtruth for MVS and SfM tasks and usually test datasets are based on the output on some state of the art algorithm, which prevents the objective evaluation of new algorithms. We propose a new method to generate artificial datasets based on true data to get realistic and measurable results. We use this method to benchmark different automatic RanSaC algorithms and find out how they compare to each other and identify each algorithm's strengths and weaknesses. This study reveals uses cases for each method and the possible tradeoffs between performance and execution time.
Fichier principal
Vignette du fichier
ransac_benchmark.pdf (8.54 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04260593 , version 1 (10-01-2024)

Identifiants

Citer

Clément Riu, Vincent Nozick, Pascal Monasse. Automatic Threshold RanSaC Algorithms for Pose Estimation Tasks. Communications in Computer and Information Science, 2023, Communications in Computer and Information Science, 1815 (1815), pp.1-20. ⟨10.1007/978-3-031-45725-8_1⟩. ⟨hal-04260593⟩
16 Consultations
12 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More