Efficient 3D Reconstruction of NeRF using Camera Pose Interpolation and Photometric Bundle Adjustment

Abstract

Figure 1: Comparison of the results of learning NeRF with the full camera poses and the interpolated-poses. The Full-Pose result uses the camera poses obtained by COLMAP with all the input images. The Interpolated-Pose result uses the poses obtained by COLMAP with several images and the interpolated poses between them as the initial poses. We apply Catmull-Rom Spline interpolation for translations and SLERP interpolation for rotations. The figure on the right shows the visualization of the camera poses using synthetic data. Interpolated-Pose generates images with almost the same quality as Full Pose.