Arnie Ramesh (@arnie_hacker)

2024-12-06 | โค๏ธ 20 | ๐Ÿ” 1


3D Gaussian splatting is definitely the future

Itโ€™s going to allow us to efficiently capture real-world in high fidelity.

This will:

  1. allow better training data & simulation/testing for self-driving & embodied robotics
  2. empower (realistic) world models like @theworldlabs for better generation (not just based on pixels but actual 3D representations)
  3. allow more immersive AR/VR experiences (imagine you can actually walk around real-buildings/cities)

Before this, the tech was:

  • photogrammetry/mesh-based methods which had weird artifacts (e.g. imagine the old Google Maps satellite view with weird shapes and flat grass etc).
  • NeRF which is incredibly inefficient and runs at like 10fps on a RTX 4080 (although much higher-resolution than previous techniques)

I understand why @janusch_patas has been so hyped about it finally haha

Excited to be working on research at the @ETH Scalable Parallel Computing Lab to make this highly efficient

Hereโ€™s a demo to give you an idea of whatโ€™s possible (and how much is left to be done!): https://www.youtube.com/watch?v=Qtjnn46rr8Q


Tags

domain-vision-3d domain-ai-ml domain-robotics domain-simulation