Question regarding per-scene hyperparameter adjustments

[deivse]

Hi, first of all, thank you for your work!

I noticed that in the official train_base.sh and train_big.sh scripts, multiple hyperparameters are adjusted on a per-scene basis. From what I understand, the results presented in the paper are also obtained with these per-scene adjustments. While that's not necessarily an issue in and of itself, I have several related concerns, and I would be grateful if you could address them:

1. To the best of my knowledge, the vast majority of 3DGS papers use a single set of hyperparameters for all scenes in a given dataset. In case there are scene-specific hyperparameter adjustments, I think it at the very least warrants explicitly mentioning that in the paper. I tried my best to find any mention of per-scene configuration differences in the paper, but wasn't able to. If this is indeed the case and I didn't simply miss something, I think this at the very least warrants creating a new revision on arXiv where this would be clearly stated in the main body of the paper. Apologies if I just missed it, and if so, please let me know which section of the paper to refer to.
2. This directly relates to the previous point - if the results in the paper are indeed obtained with per-scene hyperparameter tuning, then the comparison to other methods can only be fair as long as their hyperparameters were also tuned the same way (either identically, if applicable, or at least using the same deterministic method, e.g. grid search). Could you clarify if this is the case, or were the other methods evaluated using their default parameters?
3. In any case, it would be good to know what method you used to obtain these per-scene hyperparameter combinations (trial and error/some form of grid search, etc.) and what criteria were you tuning for.
   Thank you in advance for the response!