iComMa: Camera pose estimation by comparing and matching inverted 3D Gaussian splash

By /

Compared to the original LoFTR code, the files LoFTRsrcloftrutilscoarse_matching.py and LoFTRsrcloftrutilsfine_matching.py were modified to ensure gradient backpropagation.

Prepare data and train 3DGS models. We evaluated our approach using Blender, LLFF and 360° scene datasets provided by NeRF and Mip-NeRF 360. You can download them from the respective project pages.

Code: https://github.com/YuanSun-XJTU/iComMa
Thesis: https://arxiv.org/abs/2312.09031

Video:

Scroll to Top