Results Overfit to the training view

[fangchuan]

Hi, Thanks for your awesome work! I am trying the same sparse-view GS setting (12 ~32 input views) on my custom dataset.

After the SparseGS training procedure, I render the scene gaussians not only on the training view, but also on the interpolated trajectory (generate 25 frames between each two adjacent views).

The renderings at training viiews look reasonable:

But that at interpoloated trajectory appears significant artifacts:

sparsegs_warp20_video_rgb.mp4

I guess if the observations of those artifacts are not enough, but I have not figure out any practical solution to this problem, could help me figure out any reason or solution? @ForMyCat appreciate!

[ForMyCat]

Hi, thank you for your interest in our project.

To me it seems like even the training-view renders are having issues, where the scene almost looks semi-transparent. The first thing to verify is whether your input camera poses are accurate (90% of the time, if you see semi-transparent artifacts in training views, it is due to inaccurate input poses).

Besides that, you can try to turn down or disable the SDS loss to reduce hallucination, and turn up the depth loss to alleviate overfitting.