For a quantitative evaluation of our motion clustering results, we generated a set of dynamic scenes utilizing the Kubric toolkit. Here we show our rendering results on some of the scenes, including RGB images, depth, and motion grouping results. As evident in the video, our approach consistently produces high-quality rendering results and adeptly distinguishes the distinct motion patterns exhibited by various objects.

Our approach achieves high quality rendering of dynamic scenes and successfully discovers the moving parts. In the Editing sub-section, we provide a comparison between our segmentation mask and the RGB image, demonstrating the precision with which our mask delineates the contours of the moving parts. With the 3D part discovery, we can easily edit and render the real scene to get a brand new image.

Through the observation of the robot's actions, such as grasping a toy or inserting a peg, our method accurately identify the manipulated object, as well as the links and joints of the robot. Additionally, it is easy to replay the entire process and manipulate objects or robots at will in 3D.