3 ms·

There's some algorithms, such as Photo-SLAM and Gaussian Splatting SLAM (although far heavier and slower), that show that it is indeed possible to be able to es

There's some algorithms, such as Photo-SLAM and Gaussian Splatting SLAM (although far heavier and slower), that show that it is indeed possible to be able to estimate position and generate Gaussians in real-time. These are definitely still the early days for these techniques tho. The offline method still generates significantly higher resolution scenes of course, but as time goes on, real-time Gaussian Splatting will become more common and will be close to offline methods. This means that in the near future, we will be able to generate highly realistic scenes using Gaussian Splats on a smart edge + mobile robot in real-time and pass the splats via ROS onto another device running ROSplat (or other) and perform the visualisation there.

hirako2000 1y ago
OK. Thanks for your projections. I generate on GPU I can barely fit a large scene on 12GB of memory, and it takes many hours to produce 30k steps gaussians. I'm sure the tech will evolve, hardware too. We are just 5y away. I respect you open sourcing your work, it is innovative. Feels like a trophy splash, I suggest putting a link to something substantial, perhaps a page explaining where the tech will land and how this project fits that future, rather than a link to some LinkedIn.
- shadygm 1y ago
  Hey, I appreciate the feedback. I did not put a LinkedIn link in the post or repo, but I totally get your point about wanting something more substantial to explain the bigger picture. A lot of the motivation and reasoning behind the project is already included in the technical report PDF attached in the repository, I tried to make it as self-contained as possible for those curious about the background and use cases. That said, if I find some time, I’ll definitely consider putting together a separate page to outline where I think this kind of tool fits into the broader future of GS and robotics. Thanks again!
- somethingsome 1y ago
  Il very curious of that.. My mean training with ~25-30 high quality cameras takes around 20 minutes and some Gb of memory on a single GPU, what is the size of your large scale scenes? I see many possible optimizations to lower that number of Gb and time

hirako2000 1y ago

OK. Thanks for your projections. I generate on GPU I can barely fit a large scene on 12GB of memory, and it takes many hours to produce 30k steps gaussians. I'm sure the tech will evolve, hardware too. We are just 5y away. I respect you open sourcing your work, it is innovative. Feels like a trophy splash, I suggest putting a link to something substantial, perhaps a page explaining where the tech will land and how this project fits that future, rather than a link to some LinkedIn.

shadygm 1y ago
Hey, I appreciate the feedback. I did not put a LinkedIn link in the post or repo, but I totally get your point about wanting something more substantial to explain the bigger picture. A lot of the motivation and reasoning behind the project is already included in the technical report PDF attached in the repository, I tried to make it as self-contained as possible for those curious about the background and use cases. That said, if I find some time, I’ll definitely consider putting together a separate page to outline where I think this kind of tool fits into the broader future of GS and robotics. Thanks again!
somethingsome 1y ago
Il very curious of that.. My mean training with ~25-30 high quality cameras takes around 20 minutes and some Gb of memory on a single GPU, what is the size of your large scale scenes? I see many possible optimizations to lower that number of Gb and time