023D/4D reconstruction三维/四维重建
A dynamic scene mixes structure that stays fixed with content that moves, and a reconstruction has to keep the two apart over time.
UrbanGS separates them with 2D semantic guidance: static Gaussians stay consistent across frames, and each dynamic Gaussian carries a learnable time embedding that models its own deformation. Mirage2Matter reconstructs a scene from multi-view video with Gaussian splatting, recovers physically plausible properties with generative models, and aligns the scale to the real world, which gives an editable, physically grounded world model that trains vision-language-action policies for zero-shot transfer. With collaborators I also survey video world models.
S1
Semantic-guided Gaussian splatting for urban scenes. 2D semantic maps separate static regions from potentially dynamic objects; static Gaussians keep temporal consistency with a k-nearest-neighbour regularization for low-texture surfaces such as roads, and each dynamic Gaussian carries a learnable time embedding that models its own deformation.
S2
A physically grounded Gaussian world model built from multi-view video and off-the-shelf assets: 3D Gaussian splatting reconstructs the scene, generative models recover physically plausible properties, and a calibration target aligns the scale to the real world. Vision-language-action policies trained on its simulated data transfer zero-shot to real tasks.
S9
A survey of video world models that predict the world through video representations.
Seen through what changes, heaven and earth do not last the blink of an eye; seen through what does not change, things and I are both without end.
Su Shi 苏轼, “Former Ode on the Red Cliff” 《赤壁赋》, 1082盖，将，自，其，变，者，而，观，之，则，天，地，曾，不，能，以，一，瞬，自，其，不，变，者，而，观，之，则，物，与，我，皆，无，尽，也。
For this is what time is: the number of motion with respect to before and after.
Aristotle, Physics IV.11, 219b