It’s again an annual event of the retreat. I felt that the best part in the retreat is the discussion part. The benefits of discussion is usually not learning some concrete knowledge from someone, I really doubt that you could learn some knowledge from a talk or discussion in a detail level. What you can learn is more like a rough idea of the framework, a rough idea of the principal, a rough idea of the key points, a rough idea of the connection between this concept to that concept.
To be honest, I didn’t learn too much from the invited talks, to me the tasks that the presenter work on are too trivial. But I do enjoy the discussion with Adit. That’s the first time I understand the motivation of Gaussian splatter, the first time I realized how powerful particle based methods are . In real world, to represent an object, usually in computer vision, you take a lot of pictures of one single object, and then you use a neural network or whatever to learn a cad model like reconstruction. But how can you represent more objects? Also, what if you cut the object in half, how do you represent that object? Again, taking pictures with the cutted objects?
Gaussian splatter, introduce the particle representation for everything, use a Gaussian mixture model for a single particle and try to reconstruct the whole scene. This makes me wondering before my idea of only using Geometry, dynamics, e.g. in the rigid body setting, might be quite limited for representation. Particle representations really bring in powerful tools for vision representation.