Can we track moving objects in 3D, not just 2D video?

Existing moving object segmentation relies on 2D optical flow and treats motion as a static property across entire clips. GMOS grounds the problem in 3D space and time, analyzing each object's instantaneous motion per frame directly from RGB video. The team created GMOS-2K, a dataset of 2,210 annotated videos, and introduced MOS-I, a fine-grained evaluation protocol. The method achieves state-of-the-art results on multiple benchmarks while supporting real-time streaming inference.