PotatoPalooza
commented
2 years ago Upon further investigation it seems that this conversion is analogous to converting from OpenCV camera coordinate conventions to OpenGL's standards. The following article provides some insight into the differences between the standards. https://amytabb.com/tips/tutorials/2019/06/28/OpenCV-to-OpenGL-tutorial-essentials/. COLMAP follows the OpenCV convention while TensoRF follows the OpenGL NDC convention. This should help inform how to build the conversion script
TensoRF traces rays through a finite bounding box to render views of each scene. Because of this the arbitrary basis for the coordinate system, COLMAP often places cameras too far outside of this bounding box to create accurate reconstruction of the scene. To combat this we must convert the current camera coordinates to Normal Device Coordinates, a standard often used in game engines and 3D rendering that normalizes camera transformations to remain within a unit cube. Conversions scripts for our current sfm data must be created for the MVP