Robust semi-generalized camera pose estimation using multiple minimal solvers
| Název práce v češtině: | Robustní odhad pozice částečně obecné kamery za použití několika minimálních řešičů |
|---|---|
| Název v anglickém jazyce: | Robust semi-generalized camera pose estimation using multiple minimal solvers |
| Klíčová slova: | Minimal solvers|RANSAC|Camera geometry|Computer vision |
| Klíčová slova anglicky: | Minimal solvers|RANSAC|Camera geometry|Computer vision |
| Akademický rok vypsání: | 2020/2021 |
| Typ práce: | bakalářská práce |
| Jazyk práce: | angličtina |
| Ústav: | Katedra softwarového inženýrství (32-KSI) |
| Vedoucí / školitel: | Zuzana Kúkelová |
| Řešitel: | skrytý - zadáno a potvrzeno stud. odd. |
| Datum přihlášení: | 29.03.2021 |
| Datum zadání: | 01.04.2021 |
| Datum potvrzení stud. oddělením: | 13.04.2021 |
| Datum a čas obhajoby: | 07.09.2023 09:00 |
| Datum odevzdání elektronické podoby: | 20.07.2023 |
| Datum odevzdání tištěné podoby: | 20.07.2023 |
| Datum proběhlé obhajoby: | 07.09.2023 |
| Oponenti: | RNDr. Jan Horáček, Ph.D. |
| Zásady pro vypracování |
| Estimating the camera pose, w.r.t. a generalized camera, is an important problem in computer vision with many applications, e.g., in Structure-from-motion or in localization pipelines. Such pipelines are often based on sequences of images, where there might be a set of perspective cameras with known poses, and we are given a new image which is to be registered given the generalized camera composed of the known perspective ones.
Camera pose estimation, including both camera pose and intrinsic calibration, is usually achieved by applying a minimal solver inside a robust RANSAC-style estimator. RANSAC is a hypothesize-and-verify framework, which can handle large amounts of outliers in the input data. Using a minimal set of correspondences for the estimation is important since the processing time in RANSAC, i.e., the number of hypotheses that need to be tested exponentially grows with the size of the set used for the estimate. The predominant approach for robust camera geometry estimation is to use one fixed minimal solver in all iterations of RANSAC. Such a solver assumes a given camera and motion model, e.g., 5-DOF motion, and it samples the same type and the same number of input data (e.g., five 2D-2D point correspondences) in each iteration of RANSAC. However, it was recently shown [1] that combining multiple solvers inside RANSAC and using different combinations of correspondences may significantly improve RANSAC’s performance and the quality of the estimates produced by RANSAC. Several minimal solvers for estimating the camera pose, w.r.t. a generalized camera, were recently published [2,3]. They are usually referred to as semi-generalized pose estimation solvers. These minimal solvers differ in scene assumptions (planar, general), calibration assumptions, and especially assumptions about the number of points coming from different cameras of the generalized camera. The goals of this bachelor thesis are: 1. Study the problem of estimating the camera pose w.r.t. a generalized camera and efficiently implement different minimal solvers for this problem in C++. Here, the student should study how implementation alternatives, e.g., Sturm sequences vs. a companion matrix approach, affect the solvers’ efficiency and stability. 2. Design a RANSAC-based approach that will automatically select the “best” type of solver for the next RANSAC iteration from the considered solvers. The selection will be data-driven, and it will be based on the current state of the RANSAC (e.g., number of inliers found so far for all cameras in the generalized camera), properties of the solvers (e.g., efficiency and stability) and properties of the data (e.g., planarity). The goal is to generalize the selection criteria proposed in [1], which work for two different sources of correspondences (2D-2D and 2D-3D), to more sources of correspondences. 3. Test the designed RANSAC selection strategies and compare them with the standard RANSAC approach that is based on calling a single minimal solver, e.g., in a localization application. |
| Seznam odborné literatury |
| [1] F. Camposeco, A. Cohen, M. Pollefeys, and T. Sattler. Hybrid camera pose estimation. CVPR 2018
[2] E. Zheng and C. Wu. Structure from motion using structure-less resection. ICCV 2015 [3] S. Bhayani, T. Sattler, D. Barath, P. Beliansky, J. Heikkila, Z. Kukelova. Calibrated and Partially Calibrated Semi-Generalized Homographies. arXiv:2103.06535 |