Using the results as input of Gaussian Splatting trainning data

[DabblerGISer]

I'm currently working on combine the mast3r sfm pipeline based on what's given in kapture_mast3r_mapping.py(see branch mast3r_sfm) with gaussian splatting training.
I found that although the result of the sfm resconstruction is pretty good but the result of gaussian splatting shows the psnr is not satisfying. Anyone have some ideas?
I guess it might because of the initial distribution of the point cloud. I tried to densify the point cloud but the psnr is not improved. Could it be influenced by the initial distribution of the point cloud?

[DabblerGISer]

I tried to increase the pairs the output seems improved.