State-of-the-art in Novel View Synthesis / 3D Reconstruction — Gaussian Splatting

⏱ 19:00 Saturday, 2 September
📍 F0RTHSP4CE, Khorava St, 18 (free, donations welcome)
👅 En/Ru
🗣 @Xallt
🧑‍✈️ @cad215

We're going to glance over the latest approaches in Machine Learning for reconstructing a 3D scene simply from 2D video / photos.
Mostly the purpose of this evening will be to look at 3D Gaussian Splatting — a new method for 3D reconstruction that is (almost) state-of-the-art in terms of reconstruction quality. But that's actually not the interesting part — it can be rendered at Full HD, 100 fps (whereas NeRF approaches struggle with real-time low-resolution rendering)

Main talking points:
- Pros / cons of Gaussian Splatting against NeRFs
- How authors achieved such speed both in training and real-time rendering
- How 3D Gaussian Splatting can be used for dynamic reconstruction
- Intuitive understanding of the math for differentiable rendering (both Gaussian Splatting and NeRFs)
- Demo