How can I achieve a refined, detailed, and artifact-free Gaussian splatting model?

[lcj1984]

First of all, thank you and your team for your efforts in bringing such a great algorithm to us Gaussian splatting enthusiasts. I have successfully deployed it on my local computer and also tried training it with some datasets.

Recently, I came across a Chinese algorithm company called Pointcosm (www.pointcosm.cn). Their algorithm—I'm not sure how to describe it, maybe it's an algorithm, or perhaps some other clever techniques—produces scene models after training that are exceptionally clean and detailed. I handed my dataset over to their algorithm for training, and I was amazed—there were almost no artifacts or flaws. It was simply perfect!

However, when I use Postshot software or deploy any algorithm from GitHub to train the same dataset—even though my dataset is of very high quality and I’ve done a thorough job with aerial triangulation and photo pose alignment—the resulting scene models, while roughly acceptable at first glance, still exhibit artifacts in some areas, and there’s a hazy or foggy appearance in the air. In short, they just don’t look as clean or as refined as Pointcosm’s algorithm. They lack that level of cleanliness and detail!

Here is a link to their work showcase: https://space.bilibili.com/38468640?spm_id_from=333.788.upinfo.detail.click

I’d like to ask, how can I achieve such exquisite, detailed, and artifact-free results like Pointcosm? Did they implement some algorithmic optimizations? I’d greatly appreciate and look forward to your response. Wishing you all the best in your work and happiness every day!