To be honest, Olli, this was maybe the worst of your 3DGS. I watched your YouTube video on how you made it [ https://www.youtube.com/watch?v=1nMjOni1VDY ]. You had very expensive tools, a lidar scanner and a drone, and even a new Blender plugin, to combine two pointclouds. Still, I think your Magic Wand with 360 cameras has done so much better results. How I read this, is that those lidar scanners are not worth anything in 3DGS domain. A camera or 360 camera is better. At least now, as there is no some magical software and algorithms to use lidars in a special way. Drones are good, though. As it is impossible to get rooftops and aerial photos without them. So what I think is wrong in this scene? Many of those shipping containers are hollow. It seems to be worst when the viewing angle is something like 45 degrees from above. That must be where your Blender plugin does most of its work to combine drone photos with lidar photos. Combining models seems to be a problem with 3DGS. But also fronts and backs of the containers seem to be really bad. Only one is very good. The containers' sides, which are longer and have more clearer paintings, seem to do so much better. Even tops of the containers are way better than fronts and backs. One seems to have a big hole in it, though. Again, I put blame on the lidar and think that the drone did a decent job. I also think that this scene would have been way better by using only drone videos. Let me do here a small comparison with another similar-looking scene, Suvilahti DIY forever [ https://superspl.at/scene/9da84471 ]. One thing that should mention at first, is that it is about double in size. Therefore, one should expect it to have more details. Yet it also covers a larger area. Both have much graffiti and both are large outdoor areas. The other does not have containers. Those have long flat surfaces, that do not work well in 3DGS, but the ground is flat in both. The other have absolute no hollow parts. Even a grayish asphalt ground looks solid. It is perfect in my eyes. As far as I understood, it is a drone-only capture. Well, it seems that I couldn't say that much about differences, but I still saw it. One should definitely do a comparison of themself. A picture is worth a thousand words and a 3DGS model a million. ;) What I am more interested in, and where I am also experimenting myself, is how make the ground look solid. I remember you had some experiments with that respect also [ https://www.youtube.com/watch?v=pfCZeJHfjVs&t=340s ]. I also remember that back then you also tried to just give Postshot few videos taken with different cameras or light conditions to combine them. And it worked well. What I have found that lighting and colors do not matter that much in making of a pointcloud. As it does it with SIFT features, that rather match with edge shapes and such. Back to the ground. Using a synthetic ground and copy it 3DGS scene seems to be a fix. Yet it's only a lousy fix. There seems to be a better way. It does fixing in 3DGS training level and has nothing to do with pointclouds or post editing. And the fix is this: Film darker videos. I think that is something that has not been considered yet, and it should. One could think that darker videos only make scene darker to human eyes, so that it is only visual improvement. Yet the 3DGS algorithm seems to use pixel colors to build a scene. More darker a pixel is, the more opaque the result will be. And if you want solid ground, or a kitchen wall in that matter, you should have darker tones rather than bright colors. For example, in my small comparison above, you had an asphalt color 'adabaf' while the other one had '91897f'. There seems to be a huge difference in lightness: 55% versus 70%. Maybe therefore 3DGS algorithm considers it less dense. Or it might be that lighter video takes more random reflections that human brain just ignores but 3DGS considers part of an image. These are only my thoughts, not facts.