Dynamic 3D Gaussian Tracking for Graph-Based Neural Dynamics Modeling

Step 0: click Clear All to clear all window and reset the visualizer.

Step 1: select the object.

Step 2: select the action.

Step 3: click Run to visualize the predicted video and Splats.

Our model uses only 4 cameras for reconstructing the Gaussian Splats. Click the buttons below to change the view.

Initial state and actions

Predicted video

Original Gaussian Splats

Predicted Gaussian Splats

Notes:

Due to the computation constraints of Hugging Face Space, all results are precomputed.

Training a GS for an object takes around 30 seconds. Prediction typically takes only 1-2 seconds for each push!