Qualitative Results

We first showcase Atlas on OmniST. For each instance, we show the original video (top-left), the predicted object tracks (top-right), and the predicted state graph (bottom).

We present examples by transformation type: mixed cases involving both assembly and decomposition, followed by assembly-only and decomposition-only cases.

Qualitative Comparisons

We now compare GPT-4.1 + SAM3, TubeletGraph-Bidir, and Atlas on all examples shown above.

Failure Cases