Six-point camera pose estimation algorithm applied into Tensor-based Novel View Synthesis

Camera pose estimation is a basic step in many computer vision applications. In Tensor-based Novel View Synthesis (NVS), the simple camera pose estimation proposed by B. Rousso is limited to wider applications, for it needs a full calibration, and small rotation is assumed. To improve the synthesis performance, the author applies the simple six-point estimation algorithm to NVS, which is easy to understand and just requires a semi-calibration. The algorithm is further executed with RANSAC engine to select the real solution from multiple candidates. Experiments on both synthetic data and images are presented, and better performance and accuracy are achieved.