Diploma Thesis

Author: Bc. Lukáš Bujňák

Email: bujnak16@uniba.sk

Supervisor: Ing. Viktor Kocur, PhD.

Email: viktor.kocur@fmph.uniba.sk

Annotation Camera calibration is an essential prerequisite for many computer vision tasks. Standard methods require access to the camera and specialized patterns, or assume specific types of scenes. Neural networks can learn to calibrate cameras in general scenes. However, current methods are computationally demanding.

Aim

• Provide an overview of the topic of camera calibration using neural networks.
• Design a neural network that is computationally efficient while maintaining the best possible accuracy. When designing the network, it is especially advisable to consider knowledge distillation and transfer learning techniques.
• Evaluate the proposed networks on diverse datasets, with particular emphasis on computational efficiency and accuracy.

Literature

• Multiple View Geometry in Computer Vision 2nd edition, R. Hartley and A.Zisserman, Cambridge University Press, 2003. [ Link] [ Local ]
• Are Minimal Radial Distortion Solvers Really Necessary for Relative Pose Estimation?, V.Kocur et al., International Journal of Computer Vision, 2026. [ Link ] [ Local ]
• GeoCalib: Learning Single-Image Calibration with Geometric Optimization, A. Veicht et al., Lecture Notes in Computer Science, 2024. [ Link ] [ Local ]
• MoGe: Unlocking Accurate Monocular Geometry Estimation for Open-Domain Images with Optimal Training Supervision [ Link ] [ Local ]

Sources

• International Journal of Computer Vision [ Link ]
• European Conference on Computer Vision [ Link ]
• Conference on Computer Vision and Pattern Recognition [ Link ]

Status