Do Gaussian Scenes Contain Enough Structure for Intrinsic Segmentation?
The episode discusses a paper proposing a training-free, mask-free way to segment dynamic scenes directly from 4D Gaussian primitives. The authors argue that intrinsic signals like appearance, orientation, and motion can suffice for grouping objects if at least one modality separates them cleanly. This suggests that scene structure is latent within the Gaussian representation itself.