COLMAP camera pose reconstruction for NeRF

Question

I am currently playing around with mesh reconstruction from NeRF models and I noticed that most NeRF implementations provide a python script for the extraction of camera poses from images via COLMAP. It is my understanding that those scripts use sparse reconstruction or at least the feature extraction methods from COLMAP to determine camera poses and information. The COLMAP Tutorial for feature extraction states that intrinsic parameters can be manually specified and now I can't help but wonder if there is a more efficient way to extract camera poses from images (because the quality of the NeRF often depends heavily on accurate camera poses). Has anyone here ever tried to mess with the intrinsic parameters? Are there any more parameters to tweak to achieve better results in camera pose detection? Has anyone tried to create their own script for NeRF purposes? I am thankful for all tips and tricks or nudges in the right direction. If you think the idea to try and mess with parameters is stupid in the first place please also feel free to let me know! Also: If someone knows what parameters are really needed or most important for NeRF models (parmeters that can be reconstructed from a set of images) please let me know!

I have so far tried to understand the provided python scripts and read the COLMAP tutorial section. I have not tried to extract camera poses myself using COLMAP because at this point I am not entirely what parameters are really needed for good results in NeRF modelling.

Please excuse the tags I couldn't find any that better suited the question in the suggestions.

Answer 1

You can use other KIRI Engine app for example to get position when you take a photo. It gave me much better training results.

Answer 2

If you're working with NeRF and encountering issues where Colmap is preventing you from obtaining camera positions, I also faced a similar problem for several months as running NeRF you need the camera positions that is ditected by Colmap. I'm using TEM images to train with the NeRF model, but Colmap couldn't detect the camera positions.

I found a helpful workaround that saved me from this predicament. You can check out NerfMM. This tool works without the need for camera parameters. Give it a try; it might assist you in resolving your problem.