Skip to Main content Skip to Navigation

Robust, refined and selective matching for accurate camera pose estimation

Zhe Liu 1, 2 
2 imagine [Marne-la-Vallée]
LIGM - Laboratoire d'Informatique Gaspard-Monge, ENPC - École des Ponts ParisTech
Abstract : With the recent progress in photogrammetry, it is now possible to automatically reconstruct a model of a 3D scene from pictures or videos. The model is reconstructed in several stages. First, salient features (often points, but more generally regions) are detected in each image. Second, features that are common in images pairs are matched. Third, matched features are used to estimate the relative pose (position and orientation) of images. The global poses are then computed as well as the 3D location of these features (structure from motion). Finally, a dense 3D model can be estimated. The detection of salient features, their matching, as well as the estimation of camera poses play a crucial role in the reconstruction process. Inaccuracies or errors in these stages have a major impact on the accuracy and robustness of reconstruction for the entire scene. In this thesis, we propose better methods for feature matching and feature selection, which improve the robustness and accuracy of existing methods for camera position estimation. We first introduce a photometric pairwise constraint for feature matches (VLD), which is more reliable than geometric constraints. Then we propose a semi-local matching approach (K-VLD) using this photometric match constraint. We show that our method is very robust, not only for rigid scenes but also for non-rigid and repetitive scenes, which can improve the robustness and accuracy of pose estimation methods, such as based on RANSAC. To improve the accuracy in camera position estimation, we study the accuracy of reconstruction and pose estimation in function of the number and quality of matches. We experimentally derive a “quantity vs. quality” relation. Using this relation, we propose a method to select a subset of good matches to produce highly accurate pose estimations. We also aim at refining match position. For this, we propose an improvement of least square matching (LSM) using an irregular sampling grid and image scale exploration. We show that match refinement and match selection independently improve the reconstruction results, and when combined together, the results are further improved
Document type :
Complete list of metadata

Cited literature [87 references]  Display  Hide  Download
Contributor : ABES STAR :  Contact
Submitted on : Monday, August 24, 2015 - 3:22:12 PM
Last modification on : Saturday, January 15, 2022 - 3:57:45 AM
Long-term archiving on: : Wednesday, November 25, 2015 - 4:02:43 PM


Version validated by the jury (STAR)


  • HAL Id : tel-01186221, version 1


Zhe Liu. Robust, refined and selective matching for accurate camera pose estimation. Signal and Image Processing. Université Paris-Est, 2015. English. ⟨NNT : 2015PESC1020⟩. ⟨tel-01186221⟩



Record views


Files downloads