Master theses
Robust reconstruction of camera poses from images of a passenger car
Author
David Kramný
Year
2025
Type
Master thesis
Supervisor
RNDr. Luděk Kleprlík, Ph.D.
Reviewers
doc. Ing. Pavel Kordík, Ph.D.
Department
Summary
The thesis focuses on improving camera pose estimation from images of passenger cars for downstream novel view synthesis with the goal of avoiding reconstruction failure cases that appeared in previous works. Dataset creation methodology is proposed and Blender addon is implemented that facilitates the creation of ground truth data for synthetic datasets. Additionally, three new real datasets are provided that follow the same methodology, incorporating precise GNSS data allowing a creation of sampled datasets with pseudo ground truth camera poses. Three distinct solutions are proposed following incremental, global, and learning-based Structure-from-Motion paradigm. A custom matching strategy is utilized with masking of the dynamic parts of the scene to aid robustness. The added robustness is showcased on previously existing datasets and the newly proposed datasets. The learning based solution showing the most significant improvement.