Thank you for the nice texrecon tool. I use it in a setup where registration errors are unavoidable. I found that that sometimes rather than texrecon doing fancy per-face logic using a cost function, etc, it can be more robust just to pick for a face the view in which the face shows most "centrally". That is to say, the best view is when the ray from the camera center to the face center is closest to the ray from the camera center to the image center.
With this approach, if the camera moves in some kind of panning motion, the registration artifacts may cluster along some seams, instead of being distributed among many faces.
Thank you for the nice texrecon tool. I use it in a setup where registration errors are unavoidable. I found that that sometimes rather than texrecon doing fancy per-face logic using a cost function, etc, it can be more robust just to pick for a face the view in which the face shows most "centrally". That is to say, the best view is when the ray from the camera center to the face center is closest to the ray from the camera center to the image center.
With this approach, if the camera moves in some kind of panning motion, the registration artifacts may cluster along some seams, instead of being distributed among many faces.
This probably diverges from what you have in mind for the tool and what the publication discussed, but for the record, my implementation is here. https://github.com/oleg-alexandrov/mvs-texturing/commit/d06deda21d5601f70d85ff5af73ce862bab946f7
If there is interest in merging it in, I can spend more time doing a better integration.