Camera relocalization via differentiable rendering

Camera relocalization and frame interpolation are important topics in a wide range of research and are frequently used in fields such as artificial intelligence and robotics. These techniques are used effectively in autonomous systems by combining them with deep learning in tasks such as visual context understanding and object tracking. However, 3D imaging studies have become one of the most popular areas of study in computer graphics. In this study, an optimization-based camera relocalization approach is first proposed. Tested on different optimization methods, this approach successfully achieved camera relocalization for the target scene over a given source scene, which was possible with a distance of up to 4.24 cm and an angle of up to 3.005 degrees. In addition, within the scope of this study, a deep learning-based 3D frame interpolation study was carried out. This study, synthesized using 3D computer graphics, deep learning and convolutional neural networks, has developed a method in which the scene between two scenes whose camera position is known or estimated can be estimated. The obtained images show the performance of the method. Additionally, as a result of testing this method on the video stabilization problem, it was reported that the cutscene generation method improved the metrics on video stabilization.