Here we demonstrate 4D interpolation results across different degrees of camera motion or object motion. Given the set of dynamic 3D Gaussians from our method, we rasterize image, depth, and motion at interpolated time and view.
Depth and motion are defined at the canonical camera's coordinate. Only moving objects present non-white colors in our motion visualization.
Our model successfully interpolates images, depth, and motion, rendering high-quality outputs from predicted dynamic 3D Gaussians.
Even on scenarios with large object motion and camera motion, our model robustly outputs all estimates from
In highly challenging cases (such as minial image overlaps or extremely large motion), our model struggles to accurately determine the motion of moving objects.
However, camera motion estimation and geometry for static region remains still very robust.
The red color in the depth map denotes disocclusion where no image evidence is provided.